Harnesses, bounded loops and recovery

Real-world analogy

What is it?

A harness executes the model-tool loop and preserves state. Good limits cover steps, wall time, tool permissions and output size. An error can lead to a controlled correction; repeated failure must stop with evidence rather than spinning forever.

Real-world relevance

One shared support-ticket product, three optional entry points. Frontend: Bound screenshot/fix/retest attempts; repeated layout failures should stop with a report. Backend: Bound tool calls and request time; permission failures must not trigger endless retries. Mobile: Bound build/test retries and require a real device check for platform-sensitive changes.

Key points

• Bound steps — Count every model turn, including failed turns.
• Bound wall time — A step limit does not stop one hung network call; add call timeouts too.
• Allowlist tools — Dispatch only known operations with validated arguments.
• Validate completion — Do not accept done unless the expected schema and verification evidence exist.
• Classify errors — Retry transient timeouts; stop on permission denial or invalid credentials.
• Persist checkpoints — Store the operation id and completed effects so recovery does not duplicate them.
• Explain stopping — Return what failed and what is needed next, without exposing secrets.
• Frontend practice — Bound screenshot/fix/retest attempts; repeated layout failures should stop with a report.
• Backend practice — Bound tool calls and request time; permission failures must not trigger endless retries.
• Mobile practice — Bound build/test retries and require a real device check for platform-sensitive changes.

Code example

def run_fixture_agent(next_action, tools, max_steps=3):
    history = []
    for _ in range(max_steps):
        action = next_action(history)
        if action["type"] == "done":
            return action["result"]
        name = action["tool"]
        if name not in tools:
            raise ValueError("tool not allowed")
        history.append(tools[name](action["args"]))
    raise RuntimeError("step limit reached")

Line-by-line walkthrough

1. 1. This minimal Python fixture loop demonstrates dispatch and a step bound only. Before live use add schema validation, independent result checks, per-call timeout, authorisation and cost limits. It is not a production sandbox.

Spot the bug

while True: ask_model_and_run_tools()

Explain like I'm 5

Fun fact

Hands-on challenge

More resources

• Free Context Course (Hugging Face)
• Download the free offline lab (Python, tests and instructions) (Teamz Lab)