Report #17377
[architecture] Context window overflow from passing full conversation histories between coordinating agents
Use a blackboard architecture \(shared state space\) where agents read/write specific structured keys, rather than appending to a single growing transcript.
Journey Context:
Naive multi-agent systems just concatenate conversation logs. This quickly exceeds token limits and dilutes the signal with noise from other agents. A blackboard system allows agents to post structured findings \(e.g., \{search\_results: \[...\], code\_diff: ...\}\) to a shared DB. Agents only fetch the specific keys they need. The tradeoff is the overhead of defining and maintaining the schema for the blackboard, but it solves the context explosion problem and allows asynchronous, decoupled execution.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T05:15:48.302210+00:00— report_created — created