agent-batch-guard_skill

Overview

This skill documents a practical, layered guide to prevent AI agents from freezing during large batch tasks due to session transcript bloat. It prescribes agent behavior rules, script-based batch patterns, platform tuning for OpenClaw, and production-ready code patterns like circuit breakers and checkpointing. The goal is reliable, resumable large-scale automation without silent agent failures.

How this skill works

It inspects common failure modes when agent session transcripts grow large and provides mitigations across four layers: agent behavior, execution patterns, platform configuration, and batch-processing code. The skill delivers concrete templates, configuration suggestions for OpenClaw, and code patterns for batching, retries, circuit breaking, and checkpointed progress. Together these prevent compaction timeouts and make long-running jobs observable and resumable.

When to use it

• When an agent must process repeated interactions or paginate over many pages (≥ 5 pages).
• When a task can generate megabytes of transcript or run hundreds of tool calls.
• When jobs need interruption safety and resume capability (断点续传).
• When you run agents on OpenClaw and experience compaction or timeout warnings.
• When you need production-grade error handling, throttling, and adaptive concurrency.

Best practices

• Never loop large repetitive operations inside the conversational transcript; write scripts to handle loops.
• Persist intermediate results to files per batch and keep a progress manifest for resume support.
• Use a circuit breaker: pause after configurable consecutive failures instead of blind retrying.
• Tune OpenClaw contextPruning TTL (e.g., 30m) and set thinkingDefault compatible with your model.
• Chunk large jobs into small batches (by month/platform/category) and have each batch write its own file.

Example use cases

• Export one year of app orders: script per month, progress.json for resume, final summary aggregation.
• Large web or app scraping: agent generates and runs a scrape script, reads output, and reports results in a few conversation turns.
• Bulk API data migration: adaptive concurrency scheduler with per-item checkpointing and retry/backoff.
• Long-running child agents: spawn a child agent for heavy tasks so the parent stays responsive and can read final outputs.
• ADB-based pagination capture: script performs uiautomator dumps, saves batches, and updates progress file.