Agent Failure Modes That Grow with One Long Context
Agentic laziness, self-preferential bias and goal drift grow the longer an agent stays in one context window, as do one-shotting the whole job and declaring premature completion.
Evidence: moderate
Some agent failures are not about what is in the context but about how the agent behaves the longer it works in one window. They look like judgement problems (stopping early, grading its own work too kindly, forgetting a constraint), and adding more instructions does not fix them. They are distinct from the content-level four ways long contexts fail.
Anthropic's claude.dev post on dynamic workflows names three such failure modes of a long single context. Agentic laziness is stopping early and declaring the work done. Self-preferential bias is favouring its own findings when asked to judge them. Goal drift is losing constraints to lossy compaction. Anthropic's earlier long-running-agents post adds two from its own experiments. The agent tries to one-shot the whole job and runs out of context halfway, and a later session sees partial progress and declares completion prematurely (the instruction-side view is in premature completion). Compaction alone fixed neither. The same post lists leaving the environment broken and weak verification among the observed failures.
The remedies Anthropic reports are structural rather than prompt-based. The long-running harness limits each session to one unit of work, with progress held in files and git rather than in the window (externalised progress artefacts). The dynamic-workflows post splits work across sub-agents in patterns such as adversarial verification and loop-until-done, which counter self-preferential bias by moving judgement to a separate evaluator. Explicit completion checks, as in verify before done, target laziness and premature completion. Goal drift connects to the choice in compaction vs context reset, since a reset with a clean handoff restates the goal while repeated compaction can wear it away.
The general lesson is not to run a long task as one ever-growing conversation. Breaking it into bounded steps (for example diagnose, propose, act, verify), each starting from written state rather than a long transcript, and having verification done by a different agent or a deterministic check, addresses all five modes. The same behaviours are worth watching for in any long-lived agent session that claims success for work that never completed.
Sources
- claude.dev, 2 Jun 2026 — https://claude.dev/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code/
- Justin Young, Anthropic Engineering, 26 Nov 2025 — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents