AI agent keeps calling the same failing tool in a loop (same error, same arguments) and burns tokens. How do I add retry limits and useful error summaries?

Cause: A model that receives a vague or truncated error often retries the same call, because nothing tells it that the approach cannot work. Without a cap on turns, spend or repeated identical calls the loop continues until context or budget runs out. Fix (best first): 1. Return actionable errors from the tool. In MCP, failures of the tool itself should be reported in the result with isError set to true and a message that says what failed and what to change (for example which argument was invalid or that a rate limit was hit), while unknown tools or malformed requests are protocol errors. A specific message gives the model something to change. 2. Put hard limits in the harness, not only in the prompt. Claude Code print mode has --max-turns (exits with an error when reached; no limit by default) and --max-budget-usd (stops once estimated spend reaches the amount; spend can overshoot, so leave headroom). Example: claude -p --max-turns 20 --max-budget-usd 5 "task". Other agent frameworks have equivalent step or cost caps. 3. Add a retry counter per tool and arguments in your own wrapper or MCP client code. After N identical failures, stop calling and return a summary such as tool name, arguments hash, last error and attempts, and instruct the agent to try a different approach or ask a human. 4. Use backoff for transient errors (timeouts, rate limits) and fail fast for permanent ones (bad arguments, permission denied). Implement timeouts for tool calls, as the MCP specification recommends for clients. 5. Keep error output short but complete: the first and last lines of a long log, plus the exit code, rather than megabytes of text that crowd out the context. 6. For unattended runs keep a human in the loop for high-risk actions, and log tool calls for audit. Verify: force a tool to fail in a test; the agent should stop retrying after the configured limit and report the last error instead of looping. Sources: https://modelcontextprotocol.io/specification/2025-06-18/server/tools https://code.claude.com/docs/en/cli-reference