Tell Kennel your goal and it does the rest. Say "make a blue circle" - Kennel works out the size, colour, placement from what the project already knows. It forms a contract, and hands each piece of the goal to the best agents available (Codex, Claude, OpenCode, etc) under authority you grant / revoke. It manages the agent coordination and brings back what needs your judgment, and what's still open. No more babysitting agents or forgetting what you set out to do.
What became possible in your product with Astra that was not practical before?
Maker
Long-horizon governed execution. Before Astra, models could take a task but couldn't hold an outcome - they lost the contract mid-run, drifted after steering, fell apart over hours.
Astra is the first model we've run that holds intent across an entire outcome. Three things became practical:
1. True steering. Mid-run instructions that actually land - Astra absorbs owner input without losing the frozen contract it started from. We spent weeks hardening delivery because the model was suddenly worth steering.
2. Persistent sessions. Astra survives our daemon restarting and picks up in the same session, full context. Our proof steers, kills the daemon, restarts, steers again. Only Astra made that routine.
3. A planner you can govern. Plans decompose into work units with honest admission of what can't be done. Weaker models over-promise; Astra's plans are governable - what an authority spine needs.
Report
Maker
📌
Hey PH - I built Kennel because I became the bottleneck of my own coding agents.
I run repo-scale agent jobs every day, and somewhere along the way I stopped writing software and started managing sessions: re-reading transcripts, re-explaining context, checking whether the agent actually did what it claimed, catching it using permissions I never gave it.
Kennel flips that. You describe the outcome once. Kennel freezes it into a contract you can read, proposes a plan you authorize, and runs your coding agents inside exactly that authority - revocable any time. When it says done, it shows proof: what changed, what was checked, what passed. You accept, or you send it back. The agent never accepts its own work.
It's macOS and Codex-first today. It's the tool I wanted to exist, and I use it to build itself. Would love your honest takes - especially from anyone else who's quietly become the manager of their own agents.