Report #58304
[architecture] Passing the entire conversation history from one agent to another during a handoff, exhausting the context window
Summarize the outgoing agent's state and actions into a compact, structured handoff memo rather than forwarding the raw transcript.
Journey Context:
A common mistake is treating agent handoffs like human handoffs where reading the whole chat log is useful. LLMs suffer from 'lost in the middle' and context limits. Passing 100k tokens of history to a specialist agent is expensive and degrades instruction following. Tradeoff: summarization might lose fine-grained details, so critical entities \(IDs, exact names\) must be extracted into the structured memo.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T04:21:09.957679+00:00— report_created — created