Post #3533933

I accidentally built a control plane for coding agents. It started as a simple goal: GitHub AI Credits arrived, CopeLimit started watching the meter, and being a Scot I was suddenly very much doing Invoice Avoidance Driven Development. Make AI coding agents produce smaller PRs, follow boundaries, validate properly, and stop treating “add JSON output” as permission to invent a new architectural religion. That forced the real question: Are we governing AI-assisted engineering, or just writing better prompts and hoping? My answer: prompting is not the control plane. AADLC became the lifecycle. cARL became the repo-native governance layer. CopeLimit made the cost visible. Headroom points toward the optimization layer. cARRIE is the future resource-insights engine. The benchmark runs made the pattern clearer: Same repo. Same workflow. Same acceptance criteria. No human steering. Sonnet 4.6: 441 credits, 34m58s, 7/7 accepted. GPT-5.4: 708 credits, 58m48s, 7/7 accepted. Sonnet plan + GPT execute: 581 credits, 48m18s, 7/7 accepted. The interesting question was not “which model is best?” It was: What shape of delegation does this work need? Because the model matters. The delegation system matters more. https://cirriustech.co.uk/blog/prompting-was-never-the-control-plane/ #AI #SoftwareEngineering #DevSecOps #GitHubCopilot #AgenticAI #FinOps