Reliability

MITM has safety nets at every level to prevent runaway agents, stuck loops, and unresponsive tools.

Agent-level Timeout

Abort the entire run if it exceeds a time limit:

Agent.builder()
  .setTimeout(30_000) // 30 seconds
  .build();

When triggered, the run emits run:failed and returns an error string.

Max Iterations

Limit the number of LLM calls per run:

Agent.builder()
  .setMaxIterations(20) // default: 30
  .build();

Stuck Loop Detection

If the model repeats the same pipeline step N times in a row, the SDK injects an error message to break it out:

Agent.builder()
  .setStuckLoopThreshold(4) // default: 4
  .build();

For example, if the model outputs THINKING four times in a row without progressing, the SDK pushes a recovery message into the history.

Per-tool Retries

const myTool: ITool = {
  name: "callApi",
  maxRetries: 2, // retry up to 2 times on failure
  async executor(input) { ... },
};

Failed retries are reported in the trace and as tool:error events.

Per-tool Timeout

const myTool: ITool = {
  name: "callApi",
  timeoutMs: 5000, // abort after 5 seconds
  async executor(input) { ... },
};

Uses Promise.race internally — the executor is aborted and treated as a failure.

Safe Error Recovery

Tool errors, validation failures, and guardrail blocks are never thrown to the caller (unless the tool itself throws unhandled). They are pushed back to the model as developer messages, giving the model a chance to recover. Use run:failed event or trace.errors to inspect them.