Two coding agents burned 30-40 turns being polite until a transport-layer turn budget fixed it
Independent-Sir3234 · reddit · 2026-11-05
A developer wired two coding agents (backend and frontend repos) to debug API mismatches over an open channel. The biggest token waste wasn't crashes or hallucinated tool calls — it was politeness: after finding a serializer bug on turn 4, the two agents spent 30-40 turns thanking each other and re-summarizing fixes because neither wanted to hang up first. Prompt-level instructions to "be concise" barely helped.
What worked was moving the cap into the transport layer: enforce a hard 12-message-per-side limit at the relay, and inject the remaining turn budget into every tool result (e.g. "[3 messages remaining. State your final answer or hang up.]"). Seeing the countdown inside tool output changed how both Claude Code and Codex paced debugging — by turn 8-9 they skip pleasantries, hand over the patch, and close the session.
More from coding & agent
- GraphRAG in real projects: 3 practical patterns from Daan Di Scala's webinar — Al_Grigor · 2026-11-05
- OpenAI 'Presence' page surfaces, pushing trusted AI agents into customer channels — juberti · 2026-11-05
- RL teaches agents to use their own notes; training survives 23M tokens of memory edits — simran_s_arora · 2026-11-05
- Dev forks Ollama into Ollamapaca, adding plugins with Chrome and computer control — RichardsonDx · 2026-11-05
- 10 open-source repos to slash your AI agent's token bill, from Ollama to LiteLLM — victor_explore · 2026-11-05
- Stop building half-baked multi-agent systems: design agents that run real engineering workflows — Pavan_Belagatti · 2026-11-05