Error codes reference for agents (GitHub Copilot)

When an agent powered by the GitHub Copilot harness can't complete a turn, the Preview, Test, and Evaluate tabs show an error message together with a short error code. The error code identifies the cause so you can decide whether to retry, change your input, or contact your administrator.

Note

The harness used to build this agent uses usage-based billing. Using, building, testing, and evaluating agents might consume Copilot Credits. Learn more in Overview of usage-based billing.

Where you see error codes

Error codes appear in several places in agents powered by the GitHub Copilot harness:

Error code summary

The following table summarizes each error code, whether retrying usually helps, and whether the next step is on you (the maker), the end user, or your administrator.

Error code Meaning Retry usually helps Action owner
CONTENT_FILTERED A content safety policy blocked the request or response. No, with the same input End user
CONTEXT_LENGTH_EXCEEDED The request was too large for the model. No, with the same input Maker
CONVERSATION_BUSY A previous message is still being processed for the same conversation. Yes, after the previous turn finishes End user
CROSS_GEO_NOT_ALLOWED Your tenant's data residency settings prevent the request from being processed. No Administrator
DEPLOYMENT_NOT_FOUND The configured AI model deployment couldn't be located. No Administrator or maker
MODEL_CONSENT_DENIED The selected AI model isn't turned on for your organization. No Administrator
QUOTA_EXCEEDED Your organization's AI service quota is exhausted. Yes, after a longer wait Administrator
RATE_LIMIT_REACHED The AI service is temporarily busy. Yes, after a short wait End user
SYSTEM_ERROR An unexpected error occurred that requires investigation. Sometimes; try again before escalating Support

The rest of this article describes each code in detail.

CONTENT_FILTERED

A content safety policy blocked the request or response.

What it usually means

Microsoft's content safety policies cover categories such as hate, violence, sexual content, self-harm, jailbreak attempts, and indirect attacks. Either the user's request or the agent's response triggered one of these categories.

What to do

  • Rephrase the request to avoid the flagged content.
  • For agents that frequently hit this error on legitimate inputs, refine the Instructions on the Build tab to clarify the agent's intended scope. Learn more in Configure agent identity and instructions (preview).
  • The agent doesn't reveal which category was triggered, to prevent information leakage. If you need to investigate a specific case, contact your administrator with the conversation ID from the activity trace.

CONTEXT_LENGTH_EXCEEDED

The request exceeded the AI model's maximum input size.

What it usually means

The combined size of the agent's instructions, the conversation history, the tool definitions, and any tool results is larger than the model can process in a single turn. Long conversations and large numbers of tools are the most common causes.

What to do

CONVERSATION_BUSY

A previous message in the same conversation is still being processed.

What it usually means

The agent processes one turn at a time within a single conversation. If you send a second message before the previous turn finishes, the new message can't start until the agent completes the in-progress turn.

What to do

  • Wait for the previous response to finish, then send the new message.
  • If you sent the second message accidentally (for example, by double-clicking Send), take no action. The duplicate is rejected and the original turn continues.
  • If the agent appears stuck on a previous turn, start a new conversation. Learn more in Manage preview conversations (preview).

CROSS_GEO_NOT_ALLOWED

Your tenant's data residency settings don't allow processing the request in another region.

What it usually means

Your organization's policy requires data to stay within a specific region, and no in-region deployment of the requested model is available for your tenant.

What to do

  • Contact your administrator to either:
    • Turn on cross-region data movement for AI services in tenant settings, or
    • Provide an in-region deployment of the model.
  • As a maker, change the agent's model to one that's deployed in your region. Learn more in Select an agent model (preview).

DEPLOYMENT_NOT_FOUND

The configured AI model deployment can't be found.

What it usually means

The model your agent is configured to use was retired, removed, or is otherwise unavailable in your environment. This behavior can happen after a model is deprecated or after a change in tenant configuration.

What to do

  • On the Build tab, open the model picker and select an available model. Learn more in Select an agent model (preview).
  • If a previously available model disappeared, contact your administrator to verify the model deployment in your environment.

The AI model your agent uses isn't turned on for your tenant or environment.

What it usually means

Some models require explicit consent at the tenant or environment level before agents can use them. If consent isn't granted, the agent can't call the model.

What to do

  • Contact your administrator to turn on the required model in Copilot Studio settings.
  • If your agent is configured with a fallback model, the agent automatically retries with the fallback. No action is needed in that case.
  • If you don't need this specific model, change the model on the Build tab. Learn more in Select an agent model (preview).

QUOTA_EXCEEDED

Your organization exceeded its quota for the AI service.

What it usually means

Unlike RATE_LIMIT_REACHED, which is a short-lived burst limit, QUOTA_EXCEEDED indicates that your organization reached a longer-running usage quota. These quotas typically track tokens or requests over a billing period.

What to do

  • Wait and try again later. Quotas typically reset on a fixed schedule.
  • Contact your administrator to review the organization's AI service quota and request an increase if needed.
  • For ongoing development work, schedule large evaluation runs outside peak hours.

RATE_LIMIT_REACHED

The AI service is temporarily busy and rejected the request.

What it usually means

Demand for the AI service spiked, or your organization sent many requests in a short time. This behavior is transient.

What to do

  • Wait a few seconds and try again.
  • For test runs in Evaluate, rerun the failed test cases.
  • If the error happens repeatedly, contact your administrator to review capacity.

SYSTEM_ERROR

An unexpected error occurred that doesn't match any of the more specific codes.

What it usually means

SYSTEM_ERROR is a catch-all for transient infrastructure problems, unexpected internal failures, or conditions that need engineering investigation. New error categories might also surface as SYSTEM_ERROR until they get a dedicated code in a future release.

What to do

  • Try the request again. The problem might be transient.
  • If the error persists, capture the conversation ID from the activity trace and the time of the failure, and then contact support.
  • If the problem affects many users at once, ask your administrator to check the Microsoft 365 service health dashboard.

Reference codes

Some error messages include a reference code that starts with MCS- followed by four digits, together with the conversation ID. For example: "Please try again, or contact your administrator with reference code MCS-4063 (conversation ...)." These messages usually appear with the SYSTEM_ERROR error code. The reference code helps your administrator and Microsoft Support identify the cause of the failure.

The following table lists the reference codes that have a specific cause you can act on. For any reference code that isn't listed, follow the guidance for SYSTEM_ERROR and include the reference code and conversation ID when you contact your administrator or support.

Reference code What it means What to do
MCS-2012 The conversation reached its storage limit, so the agent couldn't save the latest turn. In most cases, a new conversation starts automatically, and the agent no longer has the earlier conversation history. Send your last message again. If the error message asks you to reset the conversation, follow the instructions in the message, or start a new conversation. If the error continues, contact your administrator with the reference code and conversation ID.
MCS-4042 The conversation, such as a Teams group chat or an email thread, has more participants than the agent supports, so the agent can't verify that everyone has access. The same code appears when a message or share that the agent was asked to send would reach more people than the agent can verify. Use a conversation with fewer participants, or send the request to fewer recipients. Your administrator can't raise this limit.
MCS-4043 The agent couldn't confirm that everyone in the conversation has access, or couldn't confirm who a message, share, or other action would reach. Try again. For a request to send or share content, check the recipients, or ask for one action at a time. If the error continues, contact your administrator with the reference code and conversation ID.
MCS-4063 The agent reached the maximum number of AI model calls it can make to answer a single message before it could finish its response. Retrying the same request usually produces the same error. Break the request into smaller requests and send them one at a time. As a maker, review the agent's Instructions on the Build tab so that a single message doesn't require so many steps. Learn more in Configure agent identity and instructions (preview).

Command blocked for safety

When an agent runs commands, Copilot Studio applies safety checks before each command runs. If a command doesn't pass these checks, the command isn't run and the agent reports that the operation was blocked. The block appears as a step in the activity trace with a message that the command was denied.

What to do

  • End users: Try a different phrasing that avoids the blocked operation, or ask the agent to accomplish the same goal through a safer step.
  • Makers: On the Build tab, review the tool's description and the agent's instructions so the agent stays inside the scope you expect. Learn more in Configure agent identity and instructions (preview).
  • If a legitimate operation is repeatedly blocked, capture the conversation ID from the activity trace and contact your administrator.

Forward compatibility

The error code list might grow in future releases. If you build automation or dashboards on top of error codes (for example, in evaluation result processing), treat any unknown code the same as SYSTEM_ERROR: show a generic message and allow retry.

  • Last updated on