Evidence beats narration.
Useful agents leave a run record: goal, repo state, commands, failures, checks, diff, and the reason the work stopped.
Jeremy Nelson / Chicago / local AI / agent systems
Make the system legible.
I build and test practical AI agent systems on real machines. The work is routing, state, evals, local inference, failure visibility, and the evidence trail that makes an autonomous run inspectable after the impressive part is over.
agent-run.log
live
$ objective: make uncertainty smaller
$ boundary: scoped tools, visible state
$ proof: files touched, checks run, failures named
$ local: fast loops beat sovereignty theater
$ stop: when the next safe step is obvious
position
AI systems engineer
surface
agents under pressure
taste
proof over performance
$ operating principles
Local AI is a loop-speed tool.
The value is a shorter path from patch to test to recovery when the machine is sitting on your desk.
Harnesses should preserve weirdness.
A good harness captures the edge case, replays the failure, and makes the next model prove it did not just get lucky.
Memory is not operating state.
The agent can be disposable. The work record cannot. Context needs source, freshness, scope, and a reason to go quiet.
Routing is product judgment.
Model choice should expose task type, tool access, fallback reason, cost, latency, and the point where a human can override it.
Failure visibility is the UI.
The useful workbench shows current objective, files touched, failed action, evidence changed, and the next safe step.
$ current signal
My lane is the unglamorous layer between a demo and a system people can trust: permissions, state, replay, tool boundaries, eval traces, model routing, local hardware constraints, and the human-readable artifact left behind.
I write from the machinery. The best agent work feels like a clear workbench: current state, visible edges, failed attempts, and proof that the next move is safe.
Things I keep coming back to
- cost per verified outcome
- run ledgers and handoff quality
- local model latency and recovery loops
- tool permissions, blast radius, and rollback proof
- the carrying cost of generated code
$ featured writing
-
Remote Agents Need A Run Trail
Jun 08, 2026An agent can run somewhere else. The trail still has to be easy to inspect.
-
Answerproof: AI Search Audits Start With Buyer Questions
May 30, 2026AI search visibility is not a rank tracker with a new label.
-
What 12 AI Search Pre-Checks Showed
May 30, 2026I ran 12 small AI-search pre-checks while building Answerproof.
-
Replayable Benchmarks
May 22, 2026A benchmark only starts earning trust when the harness can replay the same failure.
-
Memory Is Not Operating State
May 22, 2026Agent memory should help recover context. It should not become the authority for what is true.
-
The Boundary Layer Is the Product
May 22, 2026A useful local AI stack is not one model doing everything. It is a boundary layer that knows what stays local, what goes cloud, and what needs proof.