Testing the coding agent swarm.
I'm now on day three of long-horizon testing, and the current run has been going for a little over three hours, basically by itself.
You can’t see it yet, because the test is automated through Codex at the moment, but this is the furthest it has ever gotten.
At this moment, the machine has already generated more than 6,000 lines of code on the requirement. Planning and decomposition went well this time.
The assignment is to complete a white-room Go port of a widely used open-source orchestration system, written in Python, from a single requirements document, with no agentic babysitting whatsoever.
The test target is large enough to stress the agents, small enough to reason about, and disciplined enough to make the architecture legible.
It's been a fantastic run so far. I've discovered process bugs, including a couple of sneaky ones.
But the /babysitter function is working. The /rawcontext function is functional. And it's narrating about its own status like crazy.
The system has full programmatic transparency, and the whole thing is moving forward.
That’s the part that matters.
At this point, I'm less worried about writing a coherent agentic runtime that can code, so much as whether an agent system can stay oriented, inspect itself, recover from mistakes, expose its own state, and keep making progress across a project that is too large to solve in one breath.
That is the mission.
A coding agent that can run for hours, or days by itself, stay coherent, and complete monster projects reliably.
We're getting close, babies.
Love ya bunches.