Development pipeline, how?

I have been using OpenCLAW for a while, and I am a developer. So I want to know how I can use OpenCLAW in my development process. But I don’t want simple tutorials… I have a lot of experience using Claude Code and Codex. so simple tutorials are not useful for me.

I am searching the next level… something like: designing an operating system around agents: repeatable contracts, isolated execution, review gates, and durable organizational knowledge.

but I am a little lost on reading all the docs. Any source? video?

1 Like

Hey there Saul! Welcome!

Love where you’re going with this. You’re thinking beyond agents writing code and toward building an entire development system around them. That’s where things get really interesting!

Given your experience with Claude Code and Codex, I’d look at OpenClaw as the orchestration layer rather than a replacement for either. A few places to dig in and get you started:

  • Multi-agent architecture: Specialized agents for planning, implementation, and review.

  • ACP integration: Orchestrate Claude Code, Codex, and other coding agents through OpenClaw.

  • Sandboxing: Isolated environments and controlled execution.

  • Lobster: Repeatable workflows with explicit approval gates.

  • Skills: Reusable procedures and organizational knowledge.

I’d also recommend Scaling Agents on Kubernetes with acpx and ACP by OpenClaw contributor Onur Solmaz. It’s much closer to the level you’re describing! I’d start with a single pipeline: issue → investigation → implementation → testing → independent review → human approval. Get that working reliably, then expand.

Love to see more advanced discussions like this happening here! :lobster:

-Joe

This is how I run it day to day. One main agent orchestrates and owns review; workers build. Mapped to your four:

1. Repeatable contracts. Every job gets two files before any worker starts:

  • BRIEF.md: why, files to read first, the exact build steps, write scope, accuracy checks, rules (CPU/disk limits, what it may never touch), and deliverables with numeric acceptance (for example “-16 LUFS, 0 decode errors”).
  • SPAWN-TASK.md: the exact task text the worker gets.

The numbers in the brief are not typed by hand. A small facts script computes them and asserts invariants, and the brief embeds its output. If the facts are wrong, the script fails before anything is built.

2. Isolated execution. One child session per job:

sessions_spawn({
  task: <SPAWN-TASK.md>,
  model: "anthropic/claude-opus-5-5",
  completionTarget: "parent",
  cleanup: "keep",
  runTimeoutSeconds: 10800,
  cwd: "/path/to/project"
})

Write scope is one folder. Workers never message anyone, upload or publish. For repo work, worktree: true gives the child its own managed checkout.

3. Review gates. The parent never trusts the worker’s report. It re-measures the finished artifact with its own scripts and only then ships. From a batch of 4 animated video builds this week:

  • 2 passed the first review.
  • 1 died on a provider 529 with no report. I restarted it with “reuse your drafts in work/” and it passed.
  • 1 was sent back once with numbers: a sound layer sat at about -10 dB against a -2 to -6 dB target. The fix landed at -3.5 dB, and the same check confirmed it.

4. Durable knowledge. Each project has its own AGENTS.md that every worker reads, plus a QUEUE.md with one row per job (run id, session key, status, review result, delivery). When a review finds a class of bug, it becomes a rule in that AGENTS.md. Example: “never put timing on speech-to-text word times; read them from the audio”. The next worker inherits the rule.

OpenClaw gotchas I hit (2026.10.x):

  • You can’t steer a running claude-cli child. sessions_send with mode: "steer" returns guarded_injection_unsupported. Put everything in the brief, and send follow-ups after the run ends.
  • A child whose streamed output passes 20,000 JSONL lines is marked failed, even though the process may still finish. Tell workers to log to files and print short summaries.
  • nohup … & inside an exec call silently never ran. Use exec background mode and confirm the job started.
  • Keep each shell command under about 150 s. Cap parallel heavy jobs: render load starved our gateway into model timeouts.
1 Like