Report #16652
[architecture] Agent taking on the persona or instructions of a previous agent after a handoff due to mixed context histories
Strictly isolate system prompts by maintaining distinct context namespaces per agent. Only pass the structured data payload across the boundary, never the raw system instructions of the previous agent.
Journey Context:
When Agent A hands off to Agent B, if Agent B receives Agent A's system prompt in the history, Agent B may become confused about its own role \(e.g., a planner trying to write code\). Agents should only receive their own system prompt and the structured data payload from the previous step to maintain strict role adherence.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T03:14:56.144183+00:00— report_created — created