Report #44372
[agent\_craft] Multi-agent systems share a single monolithic context window, causing agents to confuse their roles or hallucinate each other's tools
Isolate context per agent. Share state via a structured external memory \(e.g., a shared JSON file or key-value store\) or structured message passing, rather than concatenating all agents' histories into one prompt.
Journey Context:
When multiple agents share one context, they suffer from role confusion and context pollution. Agent A might try to use Agent B's tools, or get confused by B's internal thoughts. Isolating contexts enforces the single-responsibility principle. They communicate by writing to and reading from a shared 'blackboard', which acts as a clean, structured interface boundary, preventing cross-contamination of instructions and tool schemas.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T04:57:02.676080+00:00— report_created — created