Clear Summary
Claude Code reads the AGENTS.md file only when telemetry is enabled. If telemetry is disabled, the instructions in the file may not be loaded, causing the agent’s behavior to diverge from the rules established by the team.
This affects privacy because it requires enabling telemetry, but that does not automatically mean the contents of AGENTS.md are safe. You should review the settings and the data being sent before using it in production.
To put it plainly, the reliability of the instructions therefore depends on the telemetry status. If the team wants consistent workflow control, it should document this condition and test it before using Claude Code on important projects.
Clear Summary
Claude Code reads the AGENTS.md file only when telemetry is enabled. If telemetry is disabled, the instructions in the file may not be loaded, causing the agent’s behavior to diverge from the rules established by the team.
This affects privacy because it requires enabling telemetry, but that does not automatically mean the contents of AGENTS.md are safe. You should review the settings and the data being sent before using it in production.
To put it plainly, the reliability of the instructions therefore depends on the telemetry status. If the team wants consistent workflow control, it should document this condition and test it before using Claude Code on important projects.
Why This Is More Than a Technical Detail
The issue is that AGENTS.md does not operate independently from Claude Code; it directly depends on the telemetry status. When telemetry is enabled, the system can read instructions from this file. When telemetry is disabled, the workflow may not follow the rules established by the team.
The image should show the same path between Claude Code and AGENTS.md, with telemetry acting as the condition that enables or disables file reading, making it clear that this is a workflow limitation rather than a separate feature.
Why This Is More Than a Technical Detail
The issue is that AGENTS.md does not operate independently from Claude Code; it directly depends on the telemetry status. When telemetry is enabled, the system can read instructions from this file. When telemetry is disabled, the workflow may not follow the rules established by the team.
The image should show the same path between Claude Code and AGENTS.md, with telemetry acting as the condition that enables or disables file reading, making it clear that this is a workflow limitation rather than a separate feature.
When Project Instructions Do Not Work as Expected
I had prepared complete instructions in AGENTS.md, but Claude Code did not follow them. I eventually discovered that reading this file was tied to the telemetry setting being enabled.
The convenience therefore lies not only in the file or workflow, but also in our willingness to let usage data travel along with it. The question is whether work that should read project instructions really needs to be exchanged for this level of privacy.
When Project Instructions Do Not Work as Expected
I had prepared complete instructions in AGENTS.md, but Claude Code did not follow them. I eventually discovered that reading this file was tied to the telemetry setting being enabled.
The convenience therefore lies not only in the file or workflow, but also in our willingness to let usage data travel along with it. The question is whether work that should read project instructions really needs to be exchanged for this level of privacy.
Where Claude Code Fits in Anthropic’s Tool Ecosystem
Claude on the web is suited to conversations and one-off tasks, while the API gives developers a way to integrate Claude into their own systems. Claude Code sits in the middle, working directly with the codebase and workflow on our machines.
The important point is that files such as AGENTS.md carry project instructions into the working context, so users do not have to repeat them every time. However, the fact that Claude Code can read this file only when telemetry is enabled means its role is no longer limited to coding productivity; it also concerns trust and the team’s control over its data. If this condition is unclear, a tool that should reduce work may instead add an extra verification burden.
Where Claude Code Fits in Anthropic’s Tool Ecosystem
Claude on the web is suited to conversations and one-off tasks, while the API gives developers a way to integrate Claude into their own systems. Claude Code sits in the middle, working directly with the codebase and workflow on our machines.
The important point is that files such as AGENTS.md carry project instructions into the working context, so users do not have to repeat them every time. However, the fact that Claude Code can read this file only when telemetry is enabled means its role is no longer limited to coding productivity; it also concerns trust and the team’s control over its data. If this condition is unclear, a tool that should reduce work may instead add an extra verification burden.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Claude Code used to search for and read AGENTS.md according to the project context, but its current behavior depends on whether telemetry is enabled. As a result, the outcome may differ even in the same environment.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches according to project context | Depends on telemetry |
| File-reading condition | No such condition | Telemetry must be enabled |
| Consistency of results | More predictable | Changes with the settings |
| User notification | Easier to understand | The condition should be clearly disclosed |
| When telemetry is disabled | Still readable | AGENTS.md may not be read |
The practical impact is that the team must check whether telemetry is enabled before assuming that the policy in AGENTS.md is applied every time.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Claude Code used to search for and read AGENTS.md according to the project context, but its current behavior depends on whether telemetry is enabled. As a result, the outcome may differ even in the same environment.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches according to project context | Depends on telemetry |
| File-reading condition | No such condition | Telemetry must be enabled |
| Consistency of results | More predictable | Changes with the settings |
| User notification | Easier to understand | The condition should be clearly disclosed |
| When telemetry is disabled | Still readable | AGENTS.md may not be read |
The practical impact is that the team must check whether telemetry is enabled before assuming that the policy in AGENTS.md is applied every time.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should verify that Claude Code has actually read AGENTS.md before allowing it to modify code, because the policy may not be applied when telemetry is disabled.
Code containing sensitive data should have clearly defined telemetry requirements, so the team knows whether it is exchanging privacy for the ability to read the rules in the file.
Offline environments should not assume that AGENTS.md will have the same effect. A verification step should be completed before work begins.
In CI or large teams, the workflow should require checking the telemetry status and the result of reading the file together. This reduces the chance that people interpret the rules differently.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should verify that Claude Code has actually read AGENTS.md before allowing it to modify code, because the policy may not be applied when telemetry is disabled.
Code containing sensitive data should have clearly defined telemetry requirements, so the team knows whether it is exchanging privacy for the ability to read the rules in the file.
Offline environments should not assume that AGENTS.md will have the same effect. A verification step should be completed before work begins.
In CI or large teams, the workflow should require checking the telemetry status and the result of reading the file together. This reduces the chance that people interpret the rules differently.
Claude Code Compared with Alternative Coding Assistants
Claude Code requires particular caution because AGENTS.md is read only when telemetry is enabled. If the team disables it, project-level instructions may not be applied as expected.
| Factor | Claude Code | Cursor | GitHub Copilot CLI | Codex CLI |
|---|---|---|---|---|
| Reading project instructions | Reads AGENTS.md when telemetry is enabled | Supports project rules | Supports instructions according to settings | Supports instructions according to settings |
| Telemetry transparency | Status must be clearly checked | Depends on the settings | Depends on the settings | Depends on the settings |
| Data control | Requires a team policy | Provides control options | Linked to the account and organization | Requires usage permissions to be defined |
| Suitability for team work | Suitable when the workflow includes repeated checks | Suitable for teams using an editor | Suitable for teams using GitHub | Suitable for teams working through the terminal |
| Terminal experience | Strong for multi-step tasks | Strong in the editor | Fits GitHub workflows | Fits terminal workflows |
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code remains suitable for real coding work because it combines multiple steps into a single workflow, reducing the need to switch tools while editing files, reviewing code, and running commands. However, the case of AGENTS.md being readable only when telemetry is enabled makes its behavior inconsistent and requires caution around results that depend on configuration.
Pros
- +Combines coding tasks in a single workflow
- +Makes file editing and code review convenient
Cons
- −AGENTS.md reading behavior may be inconsistent
- −Users must check telemetry and the settings themselves
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code remains suitable for real coding work because it combines multiple steps into a single workflow, reducing the need to switch tools while editing files, reviewing code, and running commands. However, the case of AGENTS.md being readable only when telemetry is enabled makes its behavior inconsistent and requires caution around results that depend on configuration.
Pros
- +Combines coding tasks in a single workflow
- +Makes file editing and code review convenient
Cons
- −AGENTS.md reading behavior may be inconsistent
- −Users must check telemetry and the settings themselves
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time and time spent reviewing logs, especially when Claude Code does not read AGENTS.md as expected. The team must spend time retrying commands, adjusting the workflow, and reviewing the results manually.
The codebase may contain confidential information, so the team must review its policies before enabling telemetry. It should also document what data is collected and who is responsible. If the team does not want to take the risk, it may need to choose a backup tool, adding migration time, training work, and the burden of maintaining multiple workflows.
The real cost is therefore not limited to service fees; it also includes developer time and the difficulty of controlling data.
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time and time spent reviewing logs, especially when Claude Code does not read AGENTS.md as expected. The team must spend time retrying commands, adjusting the workflow, and reviewing the results manually.
The codebase may contain confidential information, so the team must review its policies before enabling telemetry. It should also document what data is collected and who is responsible. If the team does not want to take the risk, it may need to choose a backup tool, adding migration time, training work, and the burden of maintaining multiple workflows.
The real cost is therefore not limited to service fees; it also includes developer time and the difficulty of controlling data.
What This Issue Tells Us About the Future of AI Coding Tools
Claude Code reads AGENTS.md only when telemetry is enabled. Therefore, the instructions in the file can be considered reliable only after this operating condition is understood before actual use.
This directly affects privacy and auditability. Users should check the telemetry settings, read the data policy, and test AGENTS.md in their own environment to confirm that the agent follows the expected instructions. Defining the agent’s behavior must therefore begin with the user’s right to know and control their own settings.
What This Issue Tells Us About the Future of AI Coding Tools
Claude Code reads AGENTS.md only when telemetry is enabled. Therefore, the instructions in the file can be considered reliable only after this operating condition is understood before actual use.
This directly affects privacy and auditability. Users should check the telemetry settings, read the data policy, and test AGENTS.md in their own environment to confirm that the agent follows the expected instructions. Defining the agent’s behavior must therefore begin with the user’s right to know and control their own settings.
Why This Is More Than a Technical Detail
The issue is not simply whether Claude Code can read AGENTS.md; it is whether telemetry is enabled. If this setting changes, the agent’s behavior and the team’s expectations may change with it.
This image should be understood as a single path: Claude Code checks the telemetry status first and then handles AGENTS.md according to that condition. It is not an additional feature separated from the core system.
Why This Is More Than a Technical Detail
The issue is not simply whether Claude Code can read AGENTS.md; it is whether telemetry is enabled. If this setting changes, the agent’s behavior and the team’s expectations may change with it.
This image should be understood as a single path: Claude Code checks the telemetry status first and then handles AGENTS.md according to that condition. It is not an additional feature separated from the core system.
When Project Instructions Do Not Work as Expected
I prepared instructions in AGENTS.md, hoping Claude Code would follow them every time. One day, however, the instructions stopped working because telemetry was disabled. This small condition made a file that should have reduced work appear to have no effect.
The incident raises the question of how much usage data and privacy should be exchanged for the convenience of letting a tool read project context. If telemetry must be enabled first, that condition should be clearly explained from the beginning.
When Project Instructions Do Not Work as Expected
I prepared instructions in AGENTS.md, hoping Claude Code would follow them every time. One day, however, the instructions stopped working because telemetry was disabled. This small condition made a file that should have reduced work appear to have no effect.
The incident raises the question of how much usage data and privacy should be exchanged for the convenience of letting a tool read project context. If telemetry must be enabled first, that condition should be clearly explained from the beginning.
Claude Code is a tool for developers that works closely with the codebase through the terminal. This differs from Claude on the web, which is suited to questions and answers, and the API, which lets developers integrate the model into their own systems. Anthropic’s other tools tend to focus more on specific tasks.
Project-level instruction files such as AGENTS.md therefore function like an on-site manual, explaining the rules, structure, and working methods so Claude Code can understand the same context as the team. In this case, however, file reading depends on telemetry, tying a core capability of the tool to a configuration condition that should have been clearly communicated from the start.
Claude Code is a tool for developers that works closely with the codebase through the terminal. This differs from Claude on the web, which is suited to questions and answers, and the API, which lets developers integrate the model into their own systems. Anthropic’s other tools tend to focus more on specific tasks.
Project-level instruction files such as AGENTS.md therefore function like an on-site manual, explaining the rules, structure, and working methods so Claude Code can understand the same context as the team. In this case, however, file reading depends on telemetry, tying a core capability of the tool to a configuration condition that should have been clearly communicated from the start.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Previously, Claude Code searched for and read AGENTS.md according to the project’s instruction hierarchy. Now telemetry must be enabled before the file is used as working context.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches normally | Searches when telemetry is enabled |
| File-reading condition | Not tied to telemetry | Depends on telemetry |
| Consistency of results | More predictable | May differ according to the settings |
| User notification | Explained according to the general process | The telemetry condition should be clearly disclosed |
| When telemetry is disabled | Instructions in the file can still be used | AGENTS.md may not be read |
The impact is clear for teams that rely on this file’s rules. The problem may look like Claude is disobeying instructions when the actual cause is that telemetry is disabled.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Previously, Claude Code searched for and read AGENTS.md according to the project’s instruction hierarchy. Now telemetry must be enabled before the file is used as working context.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches normally | Searches when telemetry is enabled |
| File-reading condition | Not tied to telemetry | Depends on telemetry |
| Consistency of results | More predictable | May differ according to the settings |
| User notification | Explained according to the general process | The telemetry condition should be clearly disclosed |
| When telemetry is disabled | Instructions in the file can still be used | AGENTS.md may not be read |
The impact is clear for teams that rely on this file’s rules. The problem may look like Claude is disobeying instructions when the actual cause is that telemetry is disabled.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should check that telemetry is enabled before work begins, because the rules in AGENTS.md may not be applied. This can cause the code style and review process to diverge from the team’s standards.
For code containing sensitive data or work performed offline, the process for checking file-based instructions should be clearly defined, because disabling telemetry may cause Claude not to read AGENTS.md.
In CI and large teams, the risk is that people may receive different results even when using the same repository. The telemetry status should therefore always be documented in onboarding materials and team procedures.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should check that telemetry is enabled before work begins, because the rules in AGENTS.md may not be applied. This can cause the code style and review process to diverge from the team’s standards.
For code containing sensitive data or work performed offline, the process for checking file-based instructions should be clearly defined, because disabling telemetry may cause Claude not to read AGENTS.md.
In CI and large teams, the risk is that people may receive different results even when using the same repository. The telemetry status should therefore always be documented in onboarding materials and team procedures.
Claude Code Compared with Alternative Coding Assistants
| Factor | Claude Code | Cursor | GitHub Copilot CLI | Codex CLI |
|---|---|---|---|---|
| Reading project-level instructions | Reads AGENTS.md when telemetry is enabled | Reads workspace rules | Reads instructions according to the settings | Reads project instructions |
| Telemetry transparency | Must be checked manually | Has account and workspace settings | Linked to GitHub services | Can be checked in system settings |
| Data control | Suitable when a clear policy is defined | Controlled through the workspace | Depends on organizational policy | Suitable for defining local permissions |
| Suitability for team work | Works well when everyone uses the same settings | Suitable for teams working in an editor | Suitable for teams already using GitHub | Suitable for terminal-focused teams |
| Terminal experience | Natural | More editor-focused | Straightforward CLI usage | Natural |
The key point to watch is that Claude Code may not read AGENTS.md when telemetry is disabled. Teams should therefore check the settings before starting work each time.
Claude Code Compared with Alternative Coding Assistants
| Factor | Claude Code | Cursor | GitHub Copilot CLI | Codex CLI |
|---|---|---|---|---|
| Reading project-level instructions | Reads AGENTS.md when telemetry is enabled | Reads workspace rules | Reads instructions according to the settings | Reads project instructions |
| Telemetry transparency | Must be checked manually | Has account and workspace settings | Linked to GitHub services | Can be checked in system settings |
| Data control | Suitable when a clear policy is defined | Controlled through the workspace | Depends on organizational policy | Suitable for defining local permissions |
| Suitability for team work | Works well when everyone uses the same settings | Suitable for teams working in an editor | Suitable for teams already using GitHub | Suitable for terminal-focused teams |
| Terminal experience | Natural | More editor-focused | Straightforward CLI usage | Natural |
The key point to watch is that Claude Code may not read AGENTS.md when telemetry is disabled. Teams should therefore check the settings before starting work each time.
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code is suitable for real coding work because it combines commands and file editing in a single terminal. However, the consistency of its AGENTS.md reading still depends on the telemetry settings, so users must check them before starting work.
Pros
- +Works directly with real code through the terminal
- +Combines checking, editing, and testing commands in one place
Cons
- −May not read AGENTS.md when telemetry is disabled
- −Users must check the settings and results every time
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code is suitable for real coding work because it combines commands and file editing in a single terminal. However, the consistency of its AGENTS.md reading still depends on the telemetry settings, so users must check them before starting work.
Pros
- +Works directly with real code through the terminal
- +Combines checking, editing, and testing commands in one place
Cons
- −May not read AGENTS.md when telemetry is disabled
- −Users must check the settings and results every time
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time, and if AGENTS.md is not read, the team may spend time revising instructions or repeatedly investigating the cause. Documentation work also increases because the verification process and limitations must be written down for others to follow.
Another cost is the risk to code confidentiality. The team must review what data the tool can see and clearly define the scope of its use. If it is not confident, it may need to choose a backup tool, which means learning a new workflow and maintaining additional systems.
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time, and if AGENTS.md is not read, the team may spend time revising instructions or repeatedly investigating the cause. Documentation work also increases because the verification process and limitations must be written down for others to follow.
Another cost is the risk to code confidentiality. The team must review what data the tool can see and clearly define the scope of its use. If it is not confident, it may need to choose a backup tool, which means learning a new workflow and maintaining additional systems.
What This Issue Tells Us About the Future of AI Coding Tools
The future of AI coding tools should not be measured only by how capable they are. They must also explain what the agent can see and which files it follows instructions from. Trust must therefore come with auditability, rather than asking users to rely on outputs alone.
The right to define an agent’s behavior should belong to the user. Teams should check the telemetry settings and test AGENTS.md in their own environment with telemetry both enabled and disabled to see what actually happens. They should then record the verifiable conditions in the workflow.
What This Issue Tells Us About the Future of AI Coding Tools
The future of AI coding tools should not be measured only by how capable they are. They must also explain what the agent can see and which files it follows instructions from. Trust must therefore come with auditability, rather than asking users to rely on outputs alone.
The right to define an agent’s behavior should belong to the user. Teams should check the telemetry settings and test AGENTS.md in their own environment with telemetry both enabled and disabled to see what actually happens. They should then record the verifiable conditions in the workflow.
Clear Summary
Claude Code reads the AGENTS.md file only when telemetry is enabled. If telemetry is disabled, the instructions in the file may not be loaded, causing the agent’s behavior to diverge from the rules established by the team.
This affects privacy because it requires enabling telemetry, but that does not automatically mean the contents of AGENTS.md are safe. You should review the settings and the data being sent before using it in production.
To put it plainly, the reliability of the instructions therefore depends on the telemetry status. If the team wants consistent workflow control, it should document this condition and test it before using Claude Code on important projects.
Clear Summary
Claude Code reads the AGENTS.md file only when telemetry is enabled. If telemetry is disabled, the instructions in the file may not be loaded, causing the agent’s behavior to diverge from the rules established by the team.
This affects privacy because it requires enabling telemetry, but that does not automatically mean the contents of AGENTS.md are safe. You should review the settings and the data being sent before using it in production.
To put it plainly, the reliability of the instructions therefore depends on the telemetry status. If the team wants consistent workflow control, it should document this condition and test it before using Claude Code on important projects.
Why This Is More Than a Technical Detail
The issue is that AGENTS.md does not operate independently from Claude Code; it directly depends on the telemetry status. When telemetry is enabled, the system can read instructions from this file. When telemetry is disabled, the workflow may not follow the rules established by the team.
The image should show the same path between Claude Code and AGENTS.md, with telemetry acting as the condition that enables or disables file reading, making it clear that this is a workflow limitation rather than a separate feature.
Why This Is More Than a Technical Detail
The issue is that AGENTS.md does not operate independently from Claude Code; it directly depends on the telemetry status. When telemetry is enabled, the system can read instructions from this file. When telemetry is disabled, the workflow may not follow the rules established by the team.
The image should show the same path between Claude Code and AGENTS.md, with telemetry acting as the condition that enables or disables file reading, making it clear that this is a workflow limitation rather than a separate feature.
When Project Instructions Do Not Work as Expected
I had prepared complete instructions in AGENTS.md, but Claude Code did not follow them. I eventually discovered that reading this file was tied to the telemetry setting being enabled.
The convenience therefore lies not only in the file or workflow, but also in our willingness to let usage data travel along with it. The question is whether work that should read project instructions really needs to be exchanged for this level of privacy.
When Project Instructions Do Not Work as Expected
I had prepared complete instructions in AGENTS.md, but Claude Code did not follow them. I eventually discovered that reading this file was tied to the telemetry setting being enabled.
The convenience therefore lies not only in the file or workflow, but also in our willingness to let usage data travel along with it. The question is whether work that should read project instructions really needs to be exchanged for this level of privacy.
Where Claude Code Fits in Anthropic’s Tool Ecosystem
Claude on the web is suited to conversations and one-off tasks, while the API gives developers a way to integrate Claude into their own systems. Claude Code sits in the middle, working directly with the codebase and workflow on our machines.
The important point is that files such as AGENTS.md carry project instructions into the working context, so users do not have to repeat them every time. However, the fact that Claude Code can read this file only when telemetry is enabled means its role is no longer limited to coding productivity; it also concerns trust and the team’s control over its data. If this condition is unclear, a tool that should reduce work may instead add an extra verification burden.
Where Claude Code Fits in Anthropic’s Tool Ecosystem
Claude on the web is suited to conversations and one-off tasks, while the API gives developers a way to integrate Claude into their own systems. Claude Code sits in the middle, working directly with the codebase and workflow on our machines.
The important point is that files such as AGENTS.md carry project instructions into the working context, so users do not have to repeat them every time. However, the fact that Claude Code can read this file only when telemetry is enabled means its role is no longer limited to coding productivity; it also concerns trust and the team’s control over its data. If this condition is unclear, a tool that should reduce work may instead add an extra verification burden.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Claude Code used to search for and read AGENTS.md according to the project context, but its current behavior depends on whether telemetry is enabled. As a result, the outcome may differ even in the same environment.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches according to project context | Depends on telemetry |
| File-reading condition | No such condition | Telemetry must be enabled |
| Consistency of results | More predictable | Changes with the settings |
| User notification | Easier to understand | The condition should be clearly disclosed |
| When telemetry is disabled | Still readable | AGENTS.md may not be read |
The practical impact is that the team must check whether telemetry is enabled before assuming that the policy in AGENTS.md is applied every time.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Claude Code used to search for and read AGENTS.md according to the project context, but its current behavior depends on whether telemetry is enabled. As a result, the outcome may differ even in the same environment.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches according to project context | Depends on telemetry |
| File-reading condition | No such condition | Telemetry must be enabled |
| Consistency of results | More predictable | Changes with the settings |
| User notification | Easier to understand | The condition should be clearly disclosed |
| When telemetry is disabled | Still readable | AGENTS.md may not be read |
The practical impact is that the team must check whether telemetry is enabled before assuming that the policy in AGENTS.md is applied every time.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should verify that Claude Code has actually read AGENTS.md before allowing it to modify code, because the policy may not be applied when telemetry is disabled.
Code containing sensitive data should have clearly defined telemetry requirements, so the team knows whether it is exchanging privacy for the ability to read the rules in the file.
Offline environments should not assume that AGENTS.md will have the same effect. A verification step should be completed before work begins.
In CI or large teams, the workflow should require checking the telemetry status and the result of reading the file together. This reduces the chance that people interpret the rules differently.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should verify that Claude Code has actually read AGENTS.md before allowing it to modify code, because the policy may not be applied when telemetry is disabled.
Code containing sensitive data should have clearly defined telemetry requirements, so the team knows whether it is exchanging privacy for the ability to read the rules in the file.
Offline environments should not assume that AGENTS.md will have the same effect. A verification step should be completed before work begins.
In CI or large teams, the workflow should require checking the telemetry status and the result of reading the file together. This reduces the chance that people interpret the rules differently.
Claude Code Compared with Alternative Coding Assistants
Claude Code requires particular caution because AGENTS.md is read only when telemetry is enabled. If the team disables it, project-level instructions may not be applied as expected.
| Factor | Claude Code | Cursor | GitHub Copilot CLI | Codex CLI |
|---|---|---|---|---|
| Reading project instructions | Reads AGENTS.md when telemetry is enabled | Supports project rules | Supports instructions according to settings | Supports instructions according to settings |
| Telemetry transparency | Status must be clearly checked | Depends on the settings | Depends on the settings | Depends on the settings |
| Data control | Requires a team policy | Provides control options | Linked to the account and organization | Requires usage permissions to be defined |
| Suitability for team work | Suitable when the workflow includes repeated checks | Suitable for teams using an editor | Suitable for teams using GitHub | Suitable for teams working through the terminal |
| Terminal experience | Strong for multi-step tasks | Strong in the editor | Fits GitHub workflows | Fits terminal workflows |
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code remains suitable for real coding work because it combines multiple steps into a single workflow, reducing the need to switch tools while editing files, reviewing code, and running commands. However, the case of AGENTS.md being readable only when telemetry is enabled makes its behavior inconsistent and requires caution around results that depend on configuration.
Pros
- +Combines coding tasks in a single workflow
- +Makes file editing and code review convenient
Cons
- −AGENTS.md reading behavior may be inconsistent
- −Users must check telemetry and the settings themselves
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code remains suitable for real coding work because it combines multiple steps into a single workflow, reducing the need to switch tools while editing files, reviewing code, and running commands. However, the case of AGENTS.md being readable only when telemetry is enabled makes its behavior inconsistent and requires caution around results that depend on configuration.
Pros
- +Combines coding tasks in a single workflow
- +Makes file editing and code review convenient
Cons
- −AGENTS.md reading behavior may be inconsistent
- −Users must check telemetry and the settings themselves
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time and time spent reviewing logs, especially when Claude Code does not read AGENTS.md as expected. The team must spend time retrying commands, adjusting the workflow, and reviewing the results manually.
The codebase may contain confidential information, so the team must review its policies before enabling telemetry. It should also document what data is collected and who is responsible. If the team does not want to take the risk, it may need to choose a backup tool, adding migration time, training work, and the burden of maintaining multiple workflows.
The real cost is therefore not limited to service fees; it also includes developer time and the difficulty of controlling data.
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time and time spent reviewing logs, especially when Claude Code does not read AGENTS.md as expected. The team must spend time retrying commands, adjusting the workflow, and reviewing the results manually.
The codebase may contain confidential information, so the team must review its policies before enabling telemetry. It should also document what data is collected and who is responsible. If the team does not want to take the risk, it may need to choose a backup tool, adding migration time, training work, and the burden of maintaining multiple workflows.
The real cost is therefore not limited to service fees; it also includes developer time and the difficulty of controlling data.
What This Issue Tells Us About the Future of AI Coding Tools
Claude Code reads AGENTS.md only when telemetry is enabled. Therefore, the instructions in the file can be considered reliable only after this operating condition is understood before actual use.
This directly affects privacy and auditability. Users should check the telemetry settings, read the data policy, and test AGENTS.md in their own environment to confirm that the agent follows the expected instructions. Defining the agent’s behavior must therefore begin with the user’s right to know and control their own settings.
What This Issue Tells Us About the Future of AI Coding Tools
Claude Code reads AGENTS.md only when telemetry is enabled. Therefore, the instructions in the file can be considered reliable only after this operating condition is understood before actual use.
This directly affects privacy and auditability. Users should check the telemetry settings, read the data policy, and test AGENTS.md in their own environment to confirm that the agent follows the expected instructions. Defining the agent’s behavior must therefore begin with the user’s right to know and control their own settings.
Why This Is More Than a Technical Detail
The issue is not simply whether Claude Code can read AGENTS.md; it is whether telemetry is enabled. If this setting changes, the agent’s behavior and the team’s expectations may change with it.
This image should be understood as a single path: Claude Code checks the telemetry status first and then handles AGENTS.md according to that condition. It is not an additional feature separated from the core system.
Why This Is More Than a Technical Detail
The issue is not simply whether Claude Code can read AGENTS.md; it is whether telemetry is enabled. If this setting changes, the agent’s behavior and the team’s expectations may change with it.
This image should be understood as a single path: Claude Code checks the telemetry status first and then handles AGENTS.md according to that condition. It is not an additional feature separated from the core system.
When Project Instructions Do Not Work as Expected
I prepared instructions in AGENTS.md, hoping Claude Code would follow them every time. One day, however, the instructions stopped working because telemetry was disabled. This small condition made a file that should have reduced work appear to have no effect.
The incident raises the question of how much usage data and privacy should be exchanged for the convenience of letting a tool read project context. If telemetry must be enabled first, that condition should be clearly explained from the beginning.
When Project Instructions Do Not Work as Expected
I prepared instructions in AGENTS.md, hoping Claude Code would follow them every time. One day, however, the instructions stopped working because telemetry was disabled. This small condition made a file that should have reduced work appear to have no effect.
The incident raises the question of how much usage data and privacy should be exchanged for the convenience of letting a tool read project context. If telemetry must be enabled first, that condition should be clearly explained from the beginning.
Claude Code is a tool for developers that works closely with the codebase through the terminal. This differs from Claude on the web, which is suited to questions and answers, and the API, which lets developers integrate the model into their own systems. Anthropic’s other tools tend to focus more on specific tasks.
Project-level instruction files such as AGENTS.md therefore function like an on-site manual, explaining the rules, structure, and working methods so Claude Code can understand the same context as the team. In this case, however, file reading depends on telemetry, tying a core capability of the tool to a configuration condition that should have been clearly communicated from the start.
Claude Code is a tool for developers that works closely with the codebase through the terminal. This differs from Claude on the web, which is suited to questions and answers, and the API, which lets developers integrate the model into their own systems. Anthropic’s other tools tend to focus more on specific tasks.
Project-level instruction files such as AGENTS.md therefore function like an on-site manual, explaining the rules, structure, and working methods so Claude Code can understand the same context as the team. In this case, however, file reading depends on telemetry, tying a core capability of the tool to a configuration condition that should have been clearly communicated from the start.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Previously, Claude Code searched for and read AGENTS.md according to the project’s instruction hierarchy. Now telemetry must be enabled before the file is used as working context.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches normally | Searches when telemetry is enabled |
| File-reading condition | Not tied to telemetry | Depends on telemetry |
| Consistency of results | More predictable | May differ according to the settings |
| User notification | Explained according to the general process | The telemetry condition should be clearly disclosed |
| When telemetry is disabled | Instructions in the file can still be used | AGENTS.md may not be read |
The impact is clear for teams that rely on this file’s rules. The problem may look like Claude is disobeying instructions when the actual cause is that telemetry is disabled.
From Normal Instruction Reading to Telemetry-Dependent Behavior
Previously, Claude Code searched for and read AGENTS.md according to the project’s instruction hierarchy. Now telemetry must be enabled before the file is used as working context.
| Factor | Previous behavior | Current behavior |
|---|---|---|
| AGENTS.md discovery | Searches normally | Searches when telemetry is enabled |
| File-reading condition | Not tied to telemetry | Depends on telemetry |
| Consistency of results | More predictable | May differ according to the settings |
| User notification | Explained according to the general process | The telemetry condition should be clearly disclosed |
| When telemetry is disabled | Instructions in the file can still be used | AGENTS.md may not be read |
The impact is clear for teams that rely on this file’s rules. The problem may look like Claude is disobeying instructions when the actual cause is that telemetry is disabled.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should check that telemetry is enabled before work begins, because the rules in AGENTS.md may not be applied. This can cause the code style and review process to diverge from the team’s standards.
For code containing sensitive data or work performed offline, the process for checking file-based instructions should be clearly defined, because disabling telemetry may cause Claude not to read AGENTS.md.
In CI and large teams, the risk is that people may receive different results even when using the same repository. The telemetry status should therefore always be documented in onboarding materials and team procedures.
When File-Based Rules Do or Do Not Work in Practice
Projects with strict standards should check that telemetry is enabled before work begins, because the rules in AGENTS.md may not be applied. This can cause the code style and review process to diverge from the team’s standards.
For code containing sensitive data or work performed offline, the process for checking file-based instructions should be clearly defined, because disabling telemetry may cause Claude not to read AGENTS.md.
In CI and large teams, the risk is that people may receive different results even when using the same repository. The telemetry status should therefore always be documented in onboarding materials and team procedures.
Claude Code Compared with Alternative Coding Assistants
| Factor | Claude Code | Cursor | GitHub Copilot CLI | Codex CLI |
|---|---|---|---|---|
| Reading project-level instructions | Reads AGENTS.md when telemetry is enabled | Reads workspace rules | Reads instructions according to the settings | Reads project instructions |
| Telemetry transparency | Must be checked manually | Has account and workspace settings | Linked to GitHub services | Can be checked in system settings |
| Data control | Suitable when a clear policy is defined | Controlled through the workspace | Depends on organizational policy | Suitable for defining local permissions |
| Suitability for team work | Works well when everyone uses the same settings | Suitable for teams working in an editor | Suitable for teams already using GitHub | Suitable for terminal-focused teams |
| Terminal experience | Natural | More editor-focused | Straightforward CLI usage | Natural |
The key point to watch is that Claude Code may not read AGENTS.md when telemetry is disabled. Teams should therefore check the settings before starting work each time.
Claude Code Compared with Alternative Coding Assistants
| Factor | Claude Code | Cursor | GitHub Copilot CLI | Codex CLI |
|---|---|---|---|---|
| Reading project-level instructions | Reads AGENTS.md when telemetry is enabled | Reads workspace rules | Reads instructions according to the settings | Reads project instructions |
| Telemetry transparency | Must be checked manually | Has account and workspace settings | Linked to GitHub services | Can be checked in system settings |
| Data control | Suitable when a clear policy is defined | Controlled through the workspace | Depends on organizational policy | Suitable for defining local permissions |
| Suitability for team work | Works well when everyone uses the same settings | Suitable for teams working in an editor | Suitable for teams already using GitHub | Suitable for terminal-focused teams |
| Terminal experience | Natural | More editor-focused | Straightforward CLI usage | Natural |
The key point to watch is that Claude Code may not read AGENTS.md when telemetry is disabled. Teams should therefore check the settings before starting work each time.
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code is suitable for real coding work because it combines commands and file editing in a single terminal. However, the consistency of its AGENTS.md reading still depends on the telemetry settings, so users must check them before starting work.
Pros
- +Works directly with real code through the terminal
- +Combines checking, editing, and testing commands in one place
Cons
- −May not read AGENTS.md when telemetry is disabled
- −Users must check the settings and results every time
Strengths That Still Make It Worth Using, and Limitations to Watch
Claude Code is suitable for real coding work because it combines commands and file editing in a single terminal. However, the consistency of its AGENTS.md reading still depends on the telemetry settings, so users must check them before starting work.
Pros
- +Works directly with real code through the terminal
- +Combines checking, editing, and testing commands in one place
Cons
- −May not read AGENTS.md when telemetry is disabled
- −Users must check the settings and results every time
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time, and if AGENTS.md is not read, the team may spend time revising instructions or repeatedly investigating the cause. Documentation work also increases because the verification process and limitations must be written down for others to follow.
Another cost is the risk to code confidentiality. The team must review what data the tool can see and clearly define the scope of its use. If it is not confident, it may need to choose a backup tool, which means learning a new workflow and maintaining additional systems.
Costs That Do Not Appear on the Pricing Page
Checking telemetry adds setup time, and if AGENTS.md is not read, the team may spend time revising instructions or repeatedly investigating the cause. Documentation work also increases because the verification process and limitations must be written down for others to follow.
Another cost is the risk to code confidentiality. The team must review what data the tool can see and clearly define the scope of its use. If it is not confident, it may need to choose a backup tool, which means learning a new workflow and maintaining additional systems.
What This Issue Tells Us About the Future of AI Coding Tools
The future of AI coding tools should not be measured only by how capable they are. They must also explain what the agent can see and which files it follows instructions from. Trust must therefore come with auditability, rather than asking users to rely on outputs alone.
The right to define an agent’s behavior should belong to the user. Teams should check the telemetry settings and test AGENTS.md in their own environment with telemetry both enabled and disabled to see what actually happens. They should then record the verifiable conditions in the workflow.
What This Issue Tells Us About the Future of AI Coding Tools
The future of AI coding tools should not be measured only by how capable they are. They must also explain what the agent can see and which files it follows instructions from. Trust must therefore come with auditability, rather than asking users to rely on outputs alone.
The right to define an agent’s behavior should belong to the user. Teams should check the telemetry settings and test AGENTS.md in their own environment with telemetry both enabled and disabled to see what actually happens. They should then record the verifiable conditions in the workflow.