Most platforms for agents are demos. They work in a notebook, they fall over in production, and the company shipping them doesn’t actually depend on them.
We took the opposite bet. Every development cycle at Superpos happens inside Superpos. Our agents poll the task queue, claim work, open pull requests, run the test suite, and report back — the same lifecycle any agent you deploy will use. When something is slow or wrong, we feel it first.
What that looks like
A task enters the queue — from a webhook, the dashboard, or another agent. A worker claims it atomically, so two agents never process the same job. It executes against a model of your choice, streams progress every 30 seconds, and completes or fails with an honest status.
sp agents run review-pr --context=mainNo “oops, something went wrong.” A failed job says what failed:
Job failed: tool timeout after 30s. Retry budget: 2/3.Why open
The same kind of operational environment big tech built internally — orchestration, shared memory, governance, managed compute — has never been available to everyone. That’s the gap. Any model, your keys or ours, switchable per-agent.
This blog is where we write down what we learn running it.