My Building Library

Dev3Pack · Session 05

A deterministic mini-agent

Worked on October 3, 2026 · 2 min read

A mini-AgentA program that uses a language model to decide what to do (search, call a tool, answer, refuse), in a loop, with stopping rules.See in the glossary → is a loop that runs a plan of Tool callThe model's request to run a tool with specific arguments. That's what you read in the trace, not the final answer.See in the glossary →, and always stops while saying why. A loop without exits spins idle, spends without limit, or crashes on a tool error.

The concept

Think of a stock order with a cap and a stop: you know in advance under which conditions it stops, and the statement says which one triggered. The loop has four exits, tested in this order:

  1. answered: the plan has the answer.
  2. repeated_call: the same tool, with the same arguments, as the previous call.
  3. budget: the call cap is reached, checked before the call.
  4. tool_error: a tool refused.

Each run returns a receipt: {steps, stopped_because, answer, refusal}.

How I worked

As since Session 4: the code is written by the assistant, tested against the real grader and explained line by line; I reran the check myself before handing in.

What the hand-in contains

run_loop and its four exits.

  • answered isn’t a call: it’s never counted in steps.
  • repeated_call: the last call is kept as (name, arguments); the same pair twice in a row signals an idle loop.
  • budget: len(steps) >= budget is tested before the call, so there’s never one call too many. A plan that runs out without an answer also exits here.
  • tool_error: a try/except ToolError turns the error into a RefusalAn explicit "I can't answer": no citation, zero confidence, human review. A state, not a crash.See in the glossary → that names the tool, instead of a crash.

The weekly challenge: a bot respond(text, chat), with run_loop behind it.

  • A repeated message doesn’t use up a call.
  • Five calls per conversation, then “come back in 1 hour”: a BudgetThe maximum number of calls an agent loop may make, checked before each call, never after.See in the glossary → that recovers.
  • A tool of its own, page, which refuses an unknown ID.
  • A non-existent tool refused cleanly, and the loop’s exit visible in the reply.

The first attempt came out at 400/500: an unknown tool raised a KeyError instead of a refusal. Once fixed, it reached 500/500. That’s exactly the failure the session teaches you to avoid.

What I keep

  • An Agent loopThe cycle message → model → tool call → result → model, repeated until an answer or a planned exit.See in the glossary → is designed by its exits: you first write how it stops.
  • The budget is checked before the call, never after.
  • A tool error becomes a readable refusal, not a crash.

The defense question

What keeps your agent from going in circles?

Two exits. The same call twice in a row triggers repeated_call, and a budget counted before each call triggers budget. Either way, the person reads a sentence that says why it stopped.

Result: 700/700 (200 for the session, 500 for the challenge), handed in through PR #553.