3 posts tagged with "harness"

CI pipelines were designed for humans. A human developer submits a PR, waits 20 minutes for CI, reads the results, iterates. The feedback loop is slow — but human context doesn't expire in 20 minutes, so the timing works.

Then I read Signadot's piece on CI for coding agents. It names something I'd been working around for months without clearly naming it myself. I reached out to Ani, their CTO. We talked through how their sandbox model maps onto the achieve gate in OpenSpec. Three weeks later I had a working integration.

A coding agent doesn't have that luxury. It submits a change in seconds. If it waits 20 minutes for CI feedback, it doesn't wait — it moves on. The feedback arrives in a cold context. The correction has to fight its way back in.

CI isn't wrong. It was right for what it was designed for. It's just the wrong tool for this job.

In From Cloud Native Apps to AI Native Agent Platforms: The Belts Are the Problem, I used the factory electrification story to make an argument about AI platform adoption: factory owners in the 1890s replaced steam engines with electric motors and kept the same belts, shafts, and building layouts. For thirty years, productivity barely moved. The breakthrough came when they reorganized the factory around the new technology — workflow-first, not power-transmission-first.

That post argued at the platform layer: the decisions organizations make about how to architect and run AI-native applications. The electrification analogy there was about keeping the wrong infrastructure assumptions while adopting new technology.

This post is one layer down — the development lifecycle itself. What happens to a team's coordination model when the implementation loop accelerates by an order of magnitude? The same pattern applies: if the team keeps the existing process assumptions while individual engineers adopt AI-accelerated workflows, the system neutralizes the gain.

With OpenSpec + Superpowers + Harness, I've run enough iterations to say the individual story is real. Features that used to take 2-3 days take hours. The workflow knows what done means. I'm not watching in between.

Then someone on the team wanted to use the same workflow. That's when I found out where the bottleneck had moved.

Anthropic's engineering team published a post on harness design for long-running apps. One line stopped me:

"Generators self-assess poorly — confident praise even for mediocre output."

I've been running OpenSpec's apply phase with a code review skill baked in. It runs in the same agent context as the implementation. The reviewer and the implementer are the same session. I read that line and immediately knew: that's my problem. The reviewer runs inside the same context that just wrote the code. The confidence is real. The bias is invisible.

That was the crack. This post is the fix.