harness engineering for coding agents
Harness engineering
The repo, checks, browser evidence, worktrees, traces, and review gates around a coding agent. The useful question is whether the work leaves evidence outside the agent's prose.
writing
Notes from production AI work.
Essays on harness engineering, context, evals, traces, and the parts that decide whether coding agents and AI systems hold up when real work reaches them.
current essays8 pieces
- 26-07-30AI EngineeringI ran a taste harness on a decision I had delegated7 min read
- 26-07-07AI EngineeringAgent memory needs a chain of custody11 min read
- 26-06-30AI EngineeringReference monitors for coding agents11 min read
- 26-06-24Indie HackingBuilding got cheaper. Entrepreneurship did not.29 min read
- 26-06-19AI EngineeringThe cost of sounding human6 min read
- 26-06-16AI EngineeringAgent control planes are not meta-harnesses12 min read
- 26-06-07AI EngineeringThe Operational Meta-Harness31 min read
- 26-04-30AI EngineeringHarness engineering for coding agents14 min read
threads:AI Engineering(7)Indie Hacking(1)
clusters
How the essays connect.
Each cluster starts from a concrete problem in production AI: agent reliability, governance above agent harnesses, evidence quality, and the social cost of publishing while using AI.
operational meta-harness
Operational meta-harness
The layer above native agent harnesses: routing, context, permissions, evidence, audit, memory, verification, rollback, and the decision to stop when the evidence is weak.
agent control plane vs meta-harness
Agent control plane boundary
Control planes govern agents as resources: identity, sessions, policy, MCP access, and audit. Meta-harnesses govern agent work as a process from task intake to accepted evidence.
agent memory provenance
Agent memory provenance
Long-term memory needs source identity, freshness, trust decisions, and tamper evidence. Retrieval alone cannot explain why an agent should act on remembered context.
preference elicitation for design agents
Preference elicitation
How to turn a vague rejection into bounded evidence without treating an agent's own choice as the decision-maker's preference.
AI-generated content trust
Authentic AI-era publishing
The social cost of using AI in public without becoming a polished founder account. Claims need more links to work, sources, numbers, and constraints.
indie hacking 2026 distribution
Indie hacking in 2026
The harder parts left after software gets cheaper: distribution, trust, runway, exposure, European constraints, and the cost of crossing from private work into public markets.
editorial map
What this writing is for.
Harness engineering
How coding agents become useful when the repo, tools, tests, browser, traces, and review loops are designed as one operating environment.
Production AI systems
Retrieval, workflow control, permissions, cost boundaries, observability, and the product surfaces people need when the model is wrong or unsure.
Engineering judgment
Notes on using AI tools without outsourcing taste: where automation helps, where it drifts, and how to keep software legible.