Stagent Forces Claude Code to Actually Finish the Job
World State Labs just dropped Stagent on Product Hunt, and the pitch is the kind only someone who has watched Claude Code drop a long task at 80% completion would write: drive Claude Code through long tasks it would otherwise drop.
What Stagent actually does is force a state machine on top of Claude Code. Plan, verify, review, ship - each stage runs as its own agent, and the agent that wrote the code is not the agent that grades it. This kills three failure modes anyone with Claude Code mileage has seen: self-grading, fake-passing-tests, and quietly abandoning the task. You define workflows in plain English via /stagent:create, or use one of the 14 ready-made patterns in the cookbook. There is a live browser viewer and cross-machine task resumption so a job started on the laptop can finish on the workstation.
Free, open source, MIT-style. github.com/jie-worldstatelabs/stagent.
The pattern is becoming a category. Anthropic's own Claude Cowork ships agents that critique other agents. obra/superpowers' methodology is built on staged verification. Shepherd from Stanford last week formalized typed execution traces between supervisor and worker. Stagent is the lightweight indie version of the same idea - and it ships today, not after a 56-page paper. If you live in Claude Code, install it and check back in a week.
https://stagent.worldstatelabs.com
← Back to all articles
What Stagent actually does is force a state machine on top of Claude Code. Plan, verify, review, ship - each stage runs as its own agent, and the agent that wrote the code is not the agent that grades it. This kills three failure modes anyone with Claude Code mileage has seen: self-grading, fake-passing-tests, and quietly abandoning the task. You define workflows in plain English via /stagent:create, or use one of the 14 ready-made patterns in the cookbook. There is a live browser viewer and cross-machine task resumption so a job started on the laptop can finish on the workstation.
Free, open source, MIT-style. github.com/jie-worldstatelabs/stagent.
The pattern is becoming a category. Anthropic's own Claude Cowork ships agents that critique other agents. obra/superpowers' methodology is built on staged verification. Shepherd from Stanford last week formalized typed execution traces between supervisor and worker. Stagent is the lightweight indie version of the same idea - and it ships today, not after a 56-page paper. If you live in Claude Code, install it and check back in a week.
https://stagent.worldstatelabs.com
Comments