https://michael.roth.rocks/blog/ Engineering reliable systems from unreliable AI agents. Verification surfaces, gates, and field data from autonomous development. en-us Thu, 03 Sep 2026 00:00:00 GMT https://michael.roth.rocks/blog/how-small-is-small-enough/ https://michael.roth.rocks/blog/how-small-is-small-enough/ Thu, 03 Sep 2026 00:00:00 GMT The old rule for sizing a microservice was that it should fit in one engineer's head. That turns out to be a useful shape for coding agents too. But the two bounds have stopped being the same size, and only one of them matters at 3am. https://michael.roth.rocks/blog/edit-the-json/ https://michael.roth.rocks/blog/edit-the-json/ Thu, 23 Jul 2026 00:00:00 GMT Improving a model is more like discovery than engineering. In mixlab a model is a JSON config, so trying a hunch is editing a file, not writing a training loop. Two molecule generators you can load and run today. https://michael.roth.rocks/blog/the-deterministic-ceiling/ https://michael.roth.rocks/blog/the-deterministic-ceiling/ Tue, 21 Jul 2026 00:00:00 GMT What you can verify at each pipeline stage isn't arbitrary. It's set by how constrained the artifact is, and the deterministic ceiling rises as language becomes plans, code, and tests. https://michael.roth.rocks/blog/biosurface-hackathon/ https://michael.roth.rocks/blog/biosurface-hackathon/ Wed, 15 Jul 2026 00:00:00 GMT My first hackathon. I built BioSurface, a tool that checks AI-written biology claims against the data, and a final test run caught a gap in my own pipeline. You can only verify what you can see. https://michael.roth.rocks/blog/trust-topology-trilogy/ https://michael.roth.rocks/blog/trust-topology-trilogy/ Tue, 07 Jul 2026 00:00:00 GMT I stopped asking which model is best and started asking whether I can verify what it produces. That reframing has a name, the verification surface, and it held across three very different domains: coding agents, medical imaging, and language models. https://michael.roth.rocks/blog/verification-surface/ https://michael.roth.rocks/blog/verification-surface/ Wed, 01 Jul 2026 00:00:00 GMT An AI agent can write code that compiles, passes every test, and still ships a privilege escalation. The bug isn't in the code, it's in what you checked against. A plain-language explanation of the verification surface. https://michael.roth.rocks/blog/verification-debt/ https://michael.roth.rocks/blog/verification-debt/ Wed, 24 Jun 2026 00:00:00 GMT AI agents rarely write bugs anymore; the risk is the slow rot as they pick inconsistent patterns over time. It's called verification debt, and the fix isn't more checks on the code, it's gating the plan and design before the code is ever written. https://michael.roth.rocks/blog/the-repo-is-the-memory/ https://michael.roth.rocks/blog/the-repo-is-the-memory/ Wed, 17 Jun 2026 00:00:00 GMT A coding agent rebuilds its understanding from the repo every run, so the repo is its memory. But that memory splits in two: soft living context and hard workflow state. Each needs the opposite rule. https://michael.roth.rocks/blog/measuring-engines-with-horses/ https://michael.roth.rocks/blog/measuring-engines-with-horses/ Wed, 10 Jun 2026 00:00:00 GMT An OpenAI engineering post resurfaced and everyone is arguing about one number: 3.5 pull requests per engineer per day. The argument is in the wrong domain. PRs per day is a unit we borrowed from the world that agents are replacing. https://michael.roth.rocks/blog/harness-engineering/ https://michael.roth.rocks/blog/harness-engineering/ Wed, 03 Jun 2026 00:00:00 GMT Prompt engineering and harness engineering both involve sending text to a model. The underlying skills are different. Prompt engineering is closer to writing and rhetoric. Harness engineering is closer to distributed systems design. https://michael.roth.rocks/blog/you-already-have-a-process/ https://michael.roth.rocks/blog/you-already-have-a-process/ Wed, 27 May 2026 00:00:00 GMT Teams keep asking how to catch weird behavior from production AI agents. The better question: do you have a process? Map it once with a big agent, then run targeted agents against it. https://michael.roth.rocks/blog/plan-to-throw-one-away/ https://michael.roth.rocks/blog/plan-to-throw-one-away/ Wed, 06 May 2026 00:00:00 GMT Fred Brooks said it in 1975, and it https://michael.roth.rocks/blog/the-model-is-the-pipeline/ https://michael.roth.rocks/blog/the-model-is-the-pipeline/ Tue, 05 May 2026 00:00:00 GMT Is model choice the most important thing in getting good results from agents? For me, no. The harness around the model is doing the work. https://michael.roth.rocks/blog/share-the-thing/ https://michael.roth.rocks/blog/share-the-thing/ Thu, 30 Apr 2026 00:00:00 GMT I open-sourced a small ML utility eleven days ago. More than 3,000 people are using it. Here https://michael.roth.rocks/blog/same-gates-three-models/ https://michael.roth.rocks/blog/same-gates-three-models/ Tue, 28 Apr 2026 00:00:00 GMT Same 11 gates. Three medical imaging models. Rejection rates scale with model weakness: 6.3%, 11%, 93%. Zero code changes. https://michael.roth.rocks/blog/mixlab/ https://michael.roth.rocks/blog/mixlab/ Fri, 17 Apr 2026 00:00:00 GMT Define a model in JSON, train it on your Mac, ship the same config to a cloud GPU. No code changes between platforms. Open source: mixlab. https://michael.roth.rocks/blog/delegate-outcomes-not-tasks/ https://michael.roth.rocks/blog/delegate-outcomes-not-tasks/ Thu, 16 Apr 2026 00:00:00 GMT Good managers delegate tasks. Great managers delegate outcomes. The same is true for managing AI agents. https://michael.roth.rocks/blog/the-revision-problem/ https://michael.roth.rocks/blog/the-revision-problem/ Tue, 14 Apr 2026 00:00:00 GMT Tasks that fail early and get revised have half the downstream failure rate. The most expensive thing a pipeline can do is let bad work through early. https://michael.roth.rocks/blog/questions-i-ask-every-agent/ https://michael.roth.rocks/blog/questions-i-ask-every-agent/ Thu, 09 Apr 2026 00:00:00 GMT I build by asking questions, not by issuing commands. Four questions from my Claude Code logs that make the biggest difference. https://michael.roth.rocks/blog/cognitive-debt/ https://michael.roth.rocks/blog/cognitive-debt/ Tue, 07 Apr 2026 00:00:00 GMT As people use agents to write more code, they are trading tech debt for cognitive debt. Two strategies to stay connected to code you didn https://michael.roth.rocks/blog/the-blind-spot-map/ https://michael.roth.rocks/blog/the-blind-spot-map/ Mon, 30 Mar 2026 00:00:00 GMT Low overlap doesn https://michael.roth.rocks/blog/stage-coverage-beats-gate-density/ https://michael.roth.rocks/blog/stage-coverage-beats-gate-density/ Thu, 19 Mar 2026 00:00:00 GMT Adding more reviewers doesn https://michael.roth.rocks/blog/errors-compound-forward/ https://michael.roth.rocks/blog/errors-compound-forward/ Tue, 17 Mar 2026 00:00:00 GMT Coding agents don https://michael.roth.rocks/blog/the-terraform-destroy/ https://michael.roth.rocks/blog/the-terraform-destroy/ Thu, 12 Mar 2026 00:00:00 GMT A coding agent issued a terraform destroy in dev. The fix wasn https://michael.roth.rocks/blog/the-missing-gate/ https://michael.roth.rocks/blog/the-missing-gate/ Tue, 10 Mar 2026 00:00:00 GMT A hallucinated company name in a marketing report. The root cause wasn https://michael.roth.rocks/blog/three-robot-bakers/ https://michael.roth.rocks/blog/three-robot-bakers/ Thu, 05 Mar 2026 00:00:00 GMT AI failures are inevitable. Whether they propagate is a design choice. A simple analogy for why verification checkpoints matter more than model capability.