Writing
Writing
Agentic development, the trust layer around coding agents, and what I'm learning along the way. Start with the externalized trust series.
Externalized coordination: the missing primitive for agent swarms
Runtime and orchestration are getting solved. Coordination is the layer that tells agents what stage they are in, what evidence is required, and whether they may proceed.
Externalized proof: confidence is cheap
The agent saying 'tested' is not evidence. Software factories need proof captured outside the model's narration.
Externalized process: optional becomes skipped
Agents skip optional steps. A harness turns vague instructions like 'be careful' into a path work must pass through.
Externalized memory: the work has to remember
Agents forget. Work should not. Memory is useful only when it carries source, date, certainty, and scope.
Externalized trust and the road to software factories
Software factories will not be built on agent confidence. They will be built on trust that lives outside the chat.
Seventeen comments
An agent built working backend code. Review pushed it into the right domain boundaries.
CLAUDE.md as project constitution
Our AGENTS.md hit 700 lines. I audited it and about half was stuff the model already knows.
Composing over reimplementing: how I structure Claude Code skills
I run a personal copy of every skill in my dotfiles and a team copy in the repo. This is how that works.
Ship-it: a Claude Code skill for the last mile
Lints, tests, code review, PR, ticket update. I automated the end-of-ticket sequence so I stop forgetting steps.
Five revisions of start-ticket: how a Claude Code skill learned to start work
I kept missing steps when picking up tickets. Five revisions later the skill handles it.
Building a knowledge vault that costs zero LLM tokens at retrieval
I kept losing context between Claude Code sessions. Decisions, patterns, gotchas, all gone next morning. So I built a vault.
The programming atrophy problem: AI makes you worse at coding
Three studies say AI tools make you worse at coding. I use them every day so I wanted to know how bad it actually is.
Harness engineering: the model is just the horse
Same model, same benchmark, 36-point score difference just from the setup around it. Most of my engineering work now is on that environment, not the prompts.
Why I write about agentic engineering
I'm a fullstack engineer who went deep on AI-assisted development. This is where I write down what actually works, what doesn't, and what I'm still figuring out.