What a harness is — the four layers under every coding agent — step 6 of 9

The harness's layer-2 code calls the model exactly once. If the call hits a transient 429 (rate limit) or 503 (service unavailable), it raises and the user sees a crash. Production harnesses retry transient failures with exponential backoff.

Fix call_with_retries(fn, max_attempts=3) to retry up to 3 times on transient errors (any exception with transient=True attribute). Return the first successful result. Re-raise on the final attempt failure.

Expected output:

succeeded on attempt 3