Report #75979
[frontier] Agent drifts from instructions without detection until it causes a production incident or user-visible failure
Implement a separate 'watcher' agent or validation step that evaluates the primary agent's outputs against the original constraints. The watcher receives the original system prompt and the last N outputs, but NOT the full conversation history, giving it a 'fresh eyes' perspective. Use a lighter/cheaper model for the watcher since it's doing classification, not generation. Run the watcher every N turns or on any output that modifies external state.
Journey Context:
Single-agent architectures have a fundamental blind spot: the agent cannot reliably self-assess its own drift because the drift is gradual and the agent's self-evaluation is subject to the same context effects that caused the drift. Asking 'are you still following your instructions?' to an agent that has drifted is like asking a fish about water. Production teams in 2025 are moving to dual-agent patterns: a primary agent that does the work, and a watcher agent that checks adherence. The watcher sees the original constraints but not the full conversation history, making it immune to the accumulated context that caused the primary agent to drift. The cost is roughly 2x inference, but this can be optimized: run the watcher only every 5-10 turns, only on outputs that modify state, or only when the primary agent's output patterns change \(shorter responses, fewer qualifications, different vocabulary\). The watcher pattern is becoming as standard for production agents as monitoring is for production web services.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T10:07:43.004386+00:00— report_created — created