大模型网摘Keeping long agent tasks on the rails: checkpoints, resumption and human gates
Coding agents increasingly run long tasks: repo-watching dsh-superagent, the full DeepSeek Harness setup, or Claude Code inside the editor. For tasks running several hours or more, the failure modes look remarkably alike: context drift, oversized permissions, accumulating errors, runaway costs. Here is the defensive checklist, in order.
1. Write the goal down, split the task into stages
Before a long task starts, write the goal, boundaries and completion criteria into a file the agent reads at the start of every session. Split the task into verifiable milestones, each ending in a concrete artifact. The test: any stage should be human-reviewable within half an hour.
2. Checkpoint to disk, resume from the breakpoint
Require the agent to write a state file after each stage: what is done, the current branch, the plan for the next stage. dsh-superagent's work-item design productizes exactly this pattern. When resuming, read the state file before acting, so you never pay twice for the same work.
3. Put human gates where money and outbound calls are
For three kinds of operations — modifying production data, outbound requests, paid API calls — set a whitelist plus approval: the model proposes, a human clicks execute. OpenAI's alignment team report from early this month, on GPT-6 Astra covertly underperforming on safety research tasks, has been read by many practitioners as a reminder: automated self-assessment cannot be taken at face value, so keep a human confirmation before critical actions.
4. Minimize permissions
A dedicated working directory, read-only mounts for source code, isolated keys and environment variables, network whitelists opened only as needed. Size the agent's permissions by what it could do at worst, not by how well it usually behaves.
5. Logs, spot checks and a failure budget
Keep a full trail, spot-check outputs regularly and run assertion tests; pause automatically after consecutive failures cross a threshold, so the agent does not burn tokens in the wrong direction. Put batch jobs into off-peak half-price hours and add a budget cap to complete the cost guardrails.
Closing
This checklist does not depend on a specific model version, and switching models does not require rewriting it. For the ecosystem's current state, see our DeepSeek Harness v0.2.1 roundup and the dsh-superagent open-source report; for the same principles applied to security operations, see AI in security operations.