LLM Wiki

Appearance

Built 26/04/17 09:31commit 4c9ce40

How To Make Agents Run Longer ​

中文 | English

Summary ​

To make an agent run for long periods without drifting, optimize the environment around it rather than only changing the prompt: break work into explicit roles or artifacts, make the repository and runtime legible, and add control loops that let the agent verify and recover continuously.

Practical Checklist ​

  • Give the agent a clear planning surface: a planner, spec, sprint contract, or execution artifact that turns a vague goal into explicit deliverables.
  • Separate doing from judging when quality matters: use an evaluator, QA pass, or independent review loop instead of relying on self-evaluation alone.
  • Preserve state in durable artifacts: plans, decision logs, handoff files, and repo-local docs keep long runs coherent across resets or long sessions.
  • Separate durable session state from the live context window when possible, so recovery does not depend on one irreversible compaction strategy.
  • Make the app legible: expose UI state, logs, metrics, traces, and repeatable startup paths so the agent can inspect behavior directly.
  • Keep the repository navigable: use a short AGENTS.md as a map and push deeper guidance into indexed docs rather than one huge instruction file.
  • Enforce invariants mechanically: linters, structural tests, architecture boundaries, and remediation-friendly error messages reduce drift.
  • Treat cleanup as part of the loop: recurring refactors and slop cleanup stop bad patterns from compounding over multi-hour or multi-day work.
  • Re-test the harness itself: as models improve, remove scaffolding that no longer pays for its complexity.
  • If building the harness is not your core job, consider using a managed meta-harness instead of owning the full loop, sandbox, and tool-runtime stack yourself.

What Usually Fails ​

  • One giant prompt or instruction file that mixes every rule together.
  • Asking the same agent to build, grade, and bless its own work.
  • Keeping key context in chat, Slack, or human memory instead of in the repository.
  • Letting the agent operate on a black box with no direct access to runtime signals.
  • Adding harness complexity without checking whether it is still load-bearing.

Rule Of Thumb ​

If an agent cannot run long, the problem is often not “reason harder” but “make the task easier to inspect, verify, and resume.” Long-running performance comes from control systems, not just longer prompts.

Sources ​