Select every condition that should be in the termination policy for a production AutoGen team

Multi-select·Medium·4.0 · 0·~1 min·Asked atH2o Ai LlamaIndex Vernacular Ai

Attempt it

AutoGen's **termination policy** is the rule that tells a group chat when to stop. The team keeps taking turns until at least one termination condition fires. Four kinds of conditions belong in a real production policy. A **max-message count** caps how many turns the team can take, so a stuck loop cannot run forever even if no one notices. A **token-usage budget** caps how many input plus output tokens the run can consume; that is the cost ceiling. A **text-mention sentinel** watches for a phrase like DONE or TERMINATE from a designated agent and stops as soon as it appears. This is the natural-completion path: when the work is actually finished, the team announces it. A **wall-clock timeout** stops the run if it has been going too long, even if no other condition fired. The two distractor conditions in the question, hardware temperature and pausing on the word the, are nonsense. The first is unrelated to agent behaviour; the second would fire almost every turn and stall the workflow. The real four cover cost, time, and progress, and AutoGen lets you OR them together so whichever fires first wins.

AutoGen 0.4 ships several termination conditions and the production pattern is to OR-combine four of them so the loop cannot run away along any single axis. **Max-message count** bounds turn loops. If two agents start ping-ponging without converging, this is the brake. **Token-usage budget** bounds spend per task. This is the line item that turns a bug from "oops, agent looped" into "we shipped a hundred-dollar invoice for a five-dollar task." **Text-mention sentinel** watches for a known phrase such as TERMINATE or DONE emitted by a designated agent and stops the loop the moment it appears; this is the natural-completion signal. **External timeout** is the wall-clock backstop that catches the cases where the model neither converges nor exhausts tokens but simply hangs on a long tool call. The distractors fail on first principles. Hardware temperature has no causal relationship to agent loops; if the machine is hot, you have a hardware problem, not a multi-agent problem. Pausing whenever the LLM uses the word the would fire on virtually every turn, which is not a termination policy, it is an outage. The composition rule is cost OR time OR progress. Cost cap stops runaway spend, time cap stops hangs, progress signals stop on natural completion, message cap stops thrash. With all four wired into one OR expression the team exits cleanly on whichever path it actually fails or finishes on.

TL;DR

A production AutoGen policy OR-combines a message cap, a token budget, a sentinel from a designated agent, and a wall-clock timeout so cost, hang, and natural completion all stop the loop cleanly.

Memory aid

Easy to grasp

Imagine a meeting where people keep talking forever. To end the meeting safely you set several alarms: stop after 30 turns of conversation, stop after the room has spent its budget, stop the moment the chairperson says the word DONE, and stop after one hour on the clock no matter what. Whichever alarm rings first ends the meeting. You do not pick one alarm and trust it, because each kind of meeting fails in a different way. A long argument hits the turn cap. A waste of time meeting hits the budget cap. A productive meeting reaches DONE early. A stuck on one question meeting hits the wall clock. The agent team uses the same idea: layer four alarms so no failure mode runs the bill up overnight.

Key concepts

Multi-Agent Systems LLM Frameworks Observability and Tracing

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

A multi-agent team has more ways to fail than a single-agent chain. Two agents can ping-pong forever, a worker can hang on a slow tool call, the team can converge but never announce completion, and any of these can quietly burn budget while a human is asleep. The termination policy is the contract that says "under no circumstances does this run continue past these bounds." It is one of the highest-leverage configuration choices in a production agent deployment.

This deep dive walks through the four production-grade conditions, why each maps to a distinct failure mode, why the two distractor options in the question are nonsense, how AutoGen 0.4 composes them in code, and how to layer a no-progress detector underneath as the senior move.

The four conditions that earn their slot

Max-message count is the crudest and the most universal. The team has N turns and then it stops. Set N from the longest legitimate trace you have observed plus a small margin. The cap exists for the case where every other condition has failed to fire and the loop is grinding away with no end in sight. It is a safety net, not a primary signal.

Token-usage budget is the cost ceiling. The loop's input plus output tokens are summed across turns and the run stops when the total exceeds the budget. This is the line item that converts a bug from "the agent looped overnight" into "the agent looped for ten minutes and stopped at the cap." Set the budget per task, not per session, and alert on runs that hit the cap because hitting it should be rare; if it fires every run, your budget is wrong or your agents are broken.

Text-mention sentinel is the natural-completion path. A designated agent emits a token like TERMINATE or DONE and the framework matches the token and ends the loop. The sentinel matters because without it the team has no way to say "we are done"; the only termination is then the safety nets, which is wasteful.

External timeout is the wall-clock backstop. It catches the failure mode where a tool call hangs, an LLM provider stalls, or a human in the loop step never resolves. Token spend is zero during a hang because the model is not running, so the token cap does not save you; only the wall-clock does.

Why OR-composition is the only correct logic

The distractor options and why they fail on first principles

AutoGen 0.4 code idiom and operational layering

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

AutoGen 0.4 ships MaxMessageTermination, TokenUsageTermination, TextMentionTermination, and TimeoutTermination as composable primitives joined with the | operator.
Microsoft AutoGen Studio surfaces these termination conditions as toggles in the team editor so non-engineers can configure the safety envelope.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy is the sentinel typically restricted to a single designated agent rather than any agent in the team?

If any agent can end the run by emitting the sentinel, a worker can accidentally terminate by quoting the keyword in its reasoning. Restricting the sentinel to a supervisor or critic makes the stop signal authoritative and prevents quoting as stop accidents.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Picking one condition (just max-messages, or just a sentinel) and assuming it covers every failure mode. A single condition leaves at least one runaway path open.

60 second bullets to scan on the way to the call.

List the four termination conditions a production AutoGen team should combine
Explain why the conditions are OR-combined rather than AND-combined

Primary sources. Browse if you want the original framing.