Report #62619
[architecture] Passing full chat history between agents causes context window overflow and degraded reasoning
Agents must synthesize a structured state object or summary during handoff rather than passing the raw conversational context to the next agent.
Journey Context:
A common mistake is treating multi-agent handoffs like human conversations, passing the entire messages array. This causes context length to grow quadratically, pushing out relevant system instructions and increasing cost/latency. By forcing Agent A to output only the strictly necessary structured state \(e.g., \{extracted\_data: ..., current\_step: ...\}\) for Agent B, you keep the context lean. The tradeoff is loss of nuance from the original conversation, but this forces cleaner, more robust agent boundaries.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:35:23.260593+00:00— report_created — created