Repository navigation
Home
Agent Batch Harness is a small, file-first CLI for splitting long agent work into bounded shards, recording each attempt, and resuming without reconstructing state from chat history.
Use it when work is too large for one agent session and the units can be given explicit inputs, outputs, and quality-control files. It coordinates trusted local processes; it is not a sandbox, queue service, or replacement for tests.
- Getting Started — install the CLI and run the synthetic example.
- Manifest and Status — understand the durable state contract.
- Runners and Verification — distinguish execution success from verified output.
- Recovery and Concurrency — retry safely and reclaim abandoned claims.
- Troubleshooting — diagnose common failures.
The manifest records what may run next. A runner exit code records whether the
process succeeded. Only a configured verifier may move a shard to verified.
Those facts are deliberately separate.
The project is a 0.1.0a1 reference harness. Its plain TSV, Markdown, JSON, and
log files are designed to remain inspectable and replaceable. Treat external
agent CLIs, prompts, manifests, and shell commands as trusted input.