RFC: Adapting "Effective harnesses for long-running agents" to Claude Code (and friends)

[pbrady]

Recently Anthropic released an engineering blog Effective harnesses for long-running agents with an autonomous coding agent example that mostly uses the API rather than leveraging CC's agents, skills, and slash commands.

I've been using superpowers for a couple weeks and it seems pretty close (in spirit) to what Anthropic is doing in the blog. My own use of superpowers has been on single features or goals that get broken down to 10-20 phase implementation plans rather than the 200+ feature set that Anthropic is targeting in the blog.

I'm curious what your thoughts would be on what it would take to expand the present capabilities to that scale.

The brainstorm/write-plan/subagent-driven-development cycle has been effective for me. Do you think its just a matter of launching that agent/skill cycle on each item in the JSON list (assuming we can generate a decent one)? I've not had much success with getting CC subagents to launch other subagents but perhaps a claude-code-cli skill could be used to teach CC how to launch headless CC instances using Bash?

[obra]

Indeed, as I understand it, subagents aren't capable of running subagents. I'm currently working on tools to add subagents to "superpowers for codex"...and as a side-effect...it might actually allow claude code subagents to run their own stables of subagents. (It's what you're talking about, but...with more tooling to make it run smoother and be a little easier to manage asynchronously.)

One thing we really need is a test setup to be able to test things like this ~reliably.