Report #42998
[architecture] Shared context windows in multi-agent groups lead to context poisoning where one agent's error degrades all others
Isolate agent contexts. Instead of a single shared scratchpad, use a blackboard pattern where agents publish validated, schema-contracted facts to a central state, rather than dumping raw conversational logs.
Journey Context:
When agents share a full conversation history, a hallucination by Agent A becomes part of the context for Agent B, causing cascading hallucinations. By forcing agents to publish only validated, structured state updates to a shared blackboard, you limit the blast radius of bad data. The tradeoff is the loss of subtle conversational nuance, which might be necessary for highly creative tasks but is fatal for deterministic workflows.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T02:38:42.854098+00:00— report_created — created