fix(handoff): prevent infinite router loop in HANDOFF multi-agent strategy#1279
Closed
shaileshpadave wants to merge 5 commits into
Closed
fix(handoff): prevent infinite router loop in HANDOFF multi-agent strategy#1279shaileshpadave wants to merge 5 commits into
shaileshpadave wants to merge 5 commits into
Conversation
…ategy Two changes stop the HANDOFF loop from running indefinitely: 1. AgentConfig.maxTurns default changed from 100 to 0 (unset sentinel) All compiler strategy defaults now activate correctly — HANDOFF uses agents.size() + 1 turns (one per agent plus one final DONE check), single-agent loops default to 25, swarm loops to 100. Previously the default of 100 caused every strategy to effectively ignore its natural cap. 2. Router system prompt rewritten to prevent re-delegation The router is now told each agent may be called AT MOST ONCE, that a responded agent is marked with "[agent_name]:" in conversation history, and to respond DONE once all needed agents have replied. Previously the prompt encouraged re-routing to agents that had already responded. Tests added: - testHandoffDefaultMaxTurnsIsNumAgentsPlusOne: 3-agent HANDOFF → loop < 4 - testHandoffExplicitMaxTurnsIsRespected: .maxTurns(10) → loop < 10 - testHandoffRouterPromptPreventsReDelegation: prompt contains AT MOST ONCE
…n roundtrip failure
…Controller" This reverts commit fd8f377.
Contributor
Author
|
The SDK default is 25, but users can pass any value they want: ceo = Agent( so lets not have pr for this |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What was the problem?
When running a HANDOFF multi-agent setup (like
13_hierarchical_agents.py), the router LLM would keep re-calling agents that had already responded instead of saying DONE — causing the workflow to loop indefinitely until it hit the 100-turn hard cap.Two things caused this:
The router had no memory of who already spoke. Nothing in the prompt told the LLM that an agent it already routed to had responded. So it would happily route to the same agent again and again.
The default turn limit was too high. The default max turns was 100, so even with a bad router decision the loop would grind through all 100 iterations before stopping. For a 3-agent team, the loop should stop after at most 4 turns — not 100.
How we fixed it
Router prompt rewritten — the router is now told each agent may be called at most once. Agents that have already responded leave a visible marker in the conversation history, and the router is instructed to say DONE once all needed agents have replied.
Smarter default turn limit — the default max turns for HANDOFF is now
number of agents + 1instead of 100. This means a 2-agent team caps at 3 turns, a 5-agent team at 6, and so on. ExplicitmaxTurnssettings still override this.Verified
Ran
13_hierarchical_agents.py(CEO → Engineering Lead → Backend Dev) against the fixed server — completed cleanly with no looping.