Field Notes
Articles for people shipping software with agents: concrete artifacts, source pressure, and reusable operating rules.
Concrete artifact
The argument has to touch a file, run, screenshot, log, or decision surface.
Source pressure
Outside references are used to sharpen the claim, not to decorate it.
Reusable rule
A good post ends with a test the reader can apply somewhere else.
The archive is intentionally small. The point is not to list every AI trend; it is to make each claim inspectable.
-
Field Log 5 ReferencesNote: Your Health Endpoint Silently Failing? Docker/WSL Reserve Port Ranges on Windows
A Windows health endpoint that can't bind a port reads as a phantom failure — the OS excluded the range, not a process. Here's the single-incident diagnosis and the one fix that holds.
- Field Log 4 References
TIL: A pre-submit self-check checklist cut writer revision rounds 4→2
Moving the reviewer's rejection criteria upstream into a fixed, checkable list inside the writer prompt halved revision rounds — without the open-ended self-correction that degrades output.
-
Field Log 4 ReferencesThe Agent Crossed the Line: Publishing the Guardrail Logs That Blocked It
An agent will eventually reach for rm -rf, a prod push, or ~/.ssh. Here are the real permission and scope guardrails that block it — enforced by the harness, not the model — shown as actual block records.
- Field Log 2 References
TIL: go build ./... does not refresh your daemon binary — embed changes silently no-op
We edited a //go:embed asset, ran `go build ./...`, restarted the daemon, and it served the old asset with exit 0. Here is why the build is a compile check, not a deploy artifact.
- About 4 References
Field Ops: The Missing Discipline for Production Agentic Software
A definition and a manifesto. Production agents fail silently, compound across stateful runs, and make cost a reliability variable — yet they ship without a named ops discipline. Here is the proposed one, built against the SRE precedent.
- Field Log 3 References
Review Agent Output With HTML, Don't Replace the Contract With It
Rendering agent output as HTML is the right way to get a human to look — but the render surface is not the machine-enforced gate. Confusing the two lets schema violations, hallucinated values, and even in-output exfiltration reach your publish queue.
- Field Log 5 References
Stateless MCP servers still leak shared state: the leaks we found and the contracts that fix them
An MCP server can be a stateless handler and still leak one tenant's payload into another's stream. Here is the shared-queue leak we traced, the CVE that proves it, and the one-line key change that closes it.
- Field Log 7 References
The Day Our Agent's Context Window Filled Up: A Postmortem
A long-running agent daemon degraded into incoherence as its context window saturated. Here is what broke, why monitoring missed it, and the three-contract architecture that fixed it.
- Field Log 7 References
Write the Spec Before the Prompt: A Copy-Paste Template for Spec-Driven Agent Work Orders
Prompts are one-shot utterances; specs are contracts. Here is the minimal spec template we hand to coding agents — plus the scaling rules for when one page is enough and when you need four files.
- Field Log 8 References
Valid JSON Is Not Valid Output: Wiring a Technical Contract Gate into CI for AI-Generated Changes
How to validate LLM output in CI: a six-layer contract gate — constrained decoding, agent-loop hooks, schema and policy checks, branch protection, exit-code deploy gates, and provenance — that decides what becomes project state.