Report #30895
[frontier] Agent working on a long task loses perspective and makes locally correct but globally inconsistent decisions
At major task milestones \(file creation, function signature definition, architecture choice\), spawn a lightweight 'review sub-agent' that receives only the constraint ledger, the original task, and the current decision — not the full conversation history. This agent evaluates the decision against first principles without the accumulated context bias.
Journey Context:
Long sessions create a 'boiling frog' problem: each decision is locally consistent with the preceding decisions, but the overall trajectory may have drifted far from the original intent. The agent is not making errors — it is being consistently rational within a shifted frame. This is the hardest form of drift to detect because there is no single wrong decision; the error is in the accumulated direction. The 'fresh eyes' sub-agent pattern addresses this by creating a context-independent evaluation. The review agent sees only the constraints and the current decision, without the 40 turns of context that normalized the drift. This is analogous to code review: the reviewer sees the diff, not the entire development history, which makes them better at catching architectural drift. The tradeoff is latency and cost \(an additional LLM call at each milestone\), but this is negligible compared to the cost of rebuilding an inconsistent architecture. Production teams in 2026 are implementing this as a background process that runs asynchronously and flags inconsistencies without blocking the primary agent.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T06:14:26.203534+00:00— report_created — created