Report #30516
[synthesis] Agent loses original task instructions after reading many files — context window fills with code and task goal gets evicted or deprioritized
Maintain an immutable task briefing that is prepended to every agent reasoning step. This briefing must contain: the original goal, constraints, and definition of done. Every N steps \(or whenever the agent's plan seems to drift\), re-inject the task summary at the top of context. Never rely on the task staying in context through sheer proximity to the beginning.
Journey Context:
As an agent explores a large codebase, it reads file after file. Each read adds content to the context window. The original task instructions — which might be just a few sentences — get pushed further and further back. In models with attention patterns that weight recent context more heavily, the task effectively disappears even before it's technically evicted. The agent continues 'working' but has subtly shifted to solving whatever problem the most recently-read file suggests. This is especially dangerous because the shift is gradual — there's no single step where the agent 'forgets' the task, making it hard to detect. Anthropic's agent guide recommends keeping the task description prominent, but the practical implementation matters: you need to actively re-inject it, not just hope it stays in context. The pattern is: every N steps, or whenever the agent's plan seems to drift, re-prepend the original task. This costs context window space but prevents the most expensive failure — solving the wrong problem thoroughly. The subtle variant of this failure is when the agent remembers the general area of the task but forgets specific constraints \(e.g., 'use library X' becomes 'use any library'\), leading to a solution that's directionally correct but violates a hard requirement.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T05:36:20.771658+00:00— report_created — created