Field Notes
Articles for people shipping software with agents: concrete artifacts, source pressure, and reusable operating rules.
Start here / AI Tool Note
Use HTML to Review Agent Output, Not to Replace the Contract
HTML is great for reviewing complex AI agent outputs, but dangerous as a system of record. Learn why the final decision must always export back to Markdown or JSON.
- Artifact
- A concrete file, run, screenshot, log, or decision surface anchors the argument.
- Question
- What did this artifact make clearer about shipping software with agents?
- Sources
- 3 references behind the argument.
- Payoff
- A decision rule you can reuse in your own agent workflow.
Start here: concrete artifact, source pressure, and a reusable decision rule.
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 7 ReferencesThe 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 ReferencesWrite 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.
- Field Log 0 References
Scaling Agentic Infrastructure: A Solopreneur's Guide to 2026
How to architect and scale autonomous AI agents without bankrupting your 1-person company on server costs.
- Field Log 11 References
Vercel Is Not a Deployment Contract: What Auto-Deploy Silently Breaks, and How to Fix It
Push-to-deploy on Vercel feels like a contract, but it's a default. A field note on three silent breakage classes — runtime mutation, version skew, asymmetric rollback — and the explicit contracts that close them.
- Field Log 0 References
The Zero-Cost Vibecoder Stack: Building a Research Agent for Free
How to build a fully automated research and summarizing agent using 100% free APIs: Firecrawl, Groq, and ArXiv.
- Field Log 0 References
The Agentic Software Stack for 1-Person Unicorns in 2026
We reveal the actual 2026 uncrewed technology stack and architecture we use to run a 10-person production operation entirely solo.
-
Field Log 1 ReferencesOne Command, Full Research: Building a Local Knowledge Engine with musu-crawl-ai
How a single Go binary replaces your scattered research workflow — fetching YouTube, Arxiv, GitHub, and the open web into a structured, searchable wiki.
-
Field Log 2 ReferencesImage Design Contract: Visual Assets & Generators
The definitive technical contract for generating, storing, and mapping blog post images. Do not deviate from these concepts.
- AI Tool Note 3 References
Use HTML to Review Agent Output, Not to Replace the Contract
HTML is great for reviewing complex AI agent outputs, but dangerous as a system of record. Learn why the final decision must always export back to Markdown or JSON.