Observer subagents catch Claude Code mistakes mid-task

My notes

Summary

An experimental Claude Code feature pairs every “worker” subagent (implementer or researcher) with a second “observer” subagent that watches its transcript live, tool calls and thinking blocks included, as it grows. The observer can flag mistakes and push a correction into the worker’s live session mid-task, rather than waiting for a full review after the fact.

Key Insight

• Standard adversarial review (one agent drafts, a second reviews with fresh context after the work is done) becomes continuous: the observer reads the same transcript stream the worker is generating, in real time.
• The observer is not just reading final output. It sees intermediate reasoning (“thinking blocks”) and each tool call as it happens, so it can catch a wrong turn before it compounds into wasted work.
• Correction is injected directly into the worker’s live session (not a separate report a human has to relay), closing the loop without a human in the middle.
• Acknowledged tradeoff: running an observer alongside every worker roughly doubles token spend for that task. The bet is that catching errors early nets out cheaper than a worker running long down a wrong path and needing a full redo.
• Framed as one of several “harness-level” features beyond the core model: orchestration patterns that tooling builders are layering on top of Claude Code, rather than model capability itself.