Field Notes
agentic-sdlc-patterns
Three real runs failed on the same trivial bug, and every new session wanted to try again. The fix wasn't a retry. It was a sentence in the backlog: do not pull on this thread again from this side, and here is exactly what reopens it.
18 Jul 2026
building-the-agentic-factory
The sandbox build's scariest bugs never crashed anything. A firewall rule filtering an interface that didn't exist, an adverse test that passed against a VM that wasn't there, a fix that recurred one layer up. Three containment claims, all confident, all testing nothing.
18 Jul 2026
governance-without-killing-the-joy
The same model is reachable three ways from my machine, and each way has a different answer to whether my task's content leaves the building. The gateway stopped asking which vendor an engine belongs to and started asking how it runs.
18 Jul 2026
governance-without-killing-the-joy
My eval log is a list of recorded disappointments: a fix that didn't move the number, a review panel wrongly flagging clean code three times out of four. Writing those down felt like self-sabotage. It's the only reason the eventual win means anything.
18 Jul 2026
agentic-sdlc-patterns
In two weeks across four labs, a second AI reviewer caught a fabricated example, rescoped a whole system, confirmed the same bug twice independently, and flagged a security check testing the wrong path. One thesis, five exhibits: independence is what you're paying for.
18 Jul 2026
governance-without-killing-the-joy
Go-live day is when governance gets bypassed. This site's own launch was the test: the agent's safety check flagged its own launch plan RED, the flag was honoured even though it was a false positive, and a human ran the apply.
18 Jul 2026
local-llm-lab-notes
The small local model keeps failing at the actual work, so the obvious conclusion is that it's useless. The gateway found the job it's genuinely good at: writing the structured brief a frontier model executes. The intern doesn't do the surgery. The intern preps the theatre.
18 Jul 2026
gremlins-at-work
The mirror-staleness detector ran on schedule and reported current. It was comparing one field: the latest date heading. Nine fresh entries and one stale mirror looked identical to it. The most dangerous checker is the one that runs fine and compares too little.
18 Jul 2026
governance-without-killing-the-joy
Two governance proposals were drafted, never sent, and reviewed three days later. In those three days the docs they cited had been restructured and the systems they assumed had moved. Reviewing an exception is not re-reading its words.
18 Jul 2026
local-llm-lab-notes
The symbol indexer was built in week one. It had tests, structure, and a job. Weeks later, wiring up the graph layer, I found the truth in passing: no production code had ever called it. The repo map had been silently empty on every real run.
18 Jul 2026
governance-without-killing-the-joy
Every rule in the routing policy started life as a guess with good posture. One rule carried a counter that forced a re-check after five real runs. The counter fired, the evidence came back, and the rule changed in the most boring way possible: not at all.
18 Jul 2026
agentic-sdlc-patterns
Two agent sessions closed out in parallel, both read the same last id, both wrote RUN-0023. Then it happened again with RUN-0026. Sequential ids in a multi-writer log are a collision generator by design, and the ledger sitting next door proves the alternative.
18 Jul 2026
governance-without-killing-the-joy
Two labs asked to trust cryptography that didn't exist yet, in the same week. Both times the answer was the same: the control's owner designs the lock, shape-checking is not signing, and a decision to accept a design is not the deliverable.
18 Jul 2026
personal-operating-system
My next-actions file said HIGHEST PRIORITY in capital letters, and it was pointing at work that was already finished. I couldn't see it from inside. A fresh-context model, told to trust nothing and verify the repo first, could.
15 Jul 2026
local-llm-lab-notes
The first real overnight machine-setup run failed on model names that looked completely plausible and were completely invented. The registry had tracked their checksums as unverified for weeks. Nobody had ever checked the names.
15 Jul 2026
governance-without-killing-the-joy
Three AI engines are allowed to edit my repositories. As of this week they edit a git worktree copy instead, and the human gate reviews one diff and chooses to apply it. The design work was the edges: escalations, conflicts, and the runs that must not be isolated.
15 Jul 2026
from-prompt-to-platform
A refactor of the cloud-egress gate looked tidy and passed every test. The pre-commit review board caught the one real regression: the gate had quietly learned to fail open. The reviewer was the same model that wrote the bug.
15 Jul 2026
governance-without-killing-the-joy
A privacy gate passed every test I gave it, then routed a machine-local task toward the cloud over a single underscore. The bug was not the typo. The bug was what the gate did when it met a word it didn't know.
15 Jul 2026
building-the-agentic-factory
The tempting shortcut was to treat the marketing site as low-stakes and skip the usual rigor. Instead it went through the same classification, governance, and human-apply gate as everything else the factory builds, and that turned out to matter.
3 Jul 2026
tiny-enterprise-architecture
A tool that models your whole system will eventually model itself. The day it did, it found a stale flag on its own repo, a bug that had been sleeping for months, and a promotion claim that was one missing file from being false.
3 Jul 2026
agentic-sdlc-patterns
Two agent sessions independently drafted the same wave of six blog posts at once. Nobody planned that. The fix was not to prevent it, it was to make it survivable.
3 Jul 2026
governance-without-killing-the-joy
Adding one interactive component to a static site is exactly the kind of change that tends to sprawl. This one went through an ARB review and a security pass before merging, and then became a shared pattern instead of a one-off.
3 Jul 2026
governance-without-killing-the-joy
Every automation in this factory was additive or read-only, on purpose, until the night it needed to clean up its own branches. The engineering was the easy part. The hard part was deciding what it was allowed to touch.
3 Jul 2026
mind-palace-diaries
Where things live is not the same question as what is allowed to leave. A boundary rule that only knows the first question is half a governance model.
1 Jul 2026
mind-palace-diaries
A silent guard that stops running looks identical to a guard with nothing to report. The fix is not a second checker. It is a log the first guard keeps of itself.
1 Jul 2026
mind-palace-diaries
A rule kept in your head, or even in a document, is a rule you will break. If a rule is worth having, it is worth a checker, and the first thing mine caught was me.
29 Jun 2026
mind-palace-diaries
A guard in your critical path is a guard you will learn to skip. Move the check out of band, read-only and silent until it fails, and there is nothing left to route around.
29 Jun 2026
mind-palace-diaries
A curated copy is still a copy, and copies drift. A second brain should link to its projects, not photocopy them, and copy in only the one thing it actually learns from.
28 Jun 2026
building-the-agentic-factory
I used Hekton to build Hekton. Over one weekend the scaffolding workflow built its own children, and the trail it left was the real product.
21 Jun 2026
from-prompt-to-platform
This blog is credible because it can show the machinery that made it: a team of bounded writing gremlins, and a gate that keeps all of them away from the publish button.
21 Jun 2026
governance-without-killing-the-joy
A few YAML files and a risk classifier do real governance work: they make an agent stop before anything irreversible, and turn blocked into a respected verdict.
20 Jun 2026
gremlins-at-work
Calling your automation a gremlin sounds unserious. It survives review because the name is a handle and the contract is the architecture.
19 Jun 2026
mind-palace-diaries
Context is infrastructure. The control plane is the boundary that decides what travels and what never leaves the machine, not the folder it sits in.
18 Jun 2026
building-the-agentic-factory
The fantasy is that AI does the work. The reality is a small department of named roles, each with one job and a stop condition.
17 Jun 2026
personal-operating-system
Every backlog tool I tried, I eventually stopped opening. The thing that stuck was a daily read-only brief that comes to me and never touches my task list.
15 Jun 2026
local-llm-lab-notes
The question stopped being can this Mac run a model, and became can the factory decide which worker does which job, with evidence.
13 Jun 2026
gremlins-at-work
A scaffolding agent with a silly name and a deadly serious stop condition. Quartergremlin works because it knows exactly where to quit.
11 Jun 2026
governance-without-killing-the-joy
The morning a commit ate my commit. Git stopped being a filing cabinet and became the audit trail, the handoff boundary, and the review checkpoint.
9 Jun 2026
building-the-agentic-factory
Vibe coding works until you try to trust it twice. The fix was not a better prompt. It was an operating model.
7 Jun 2026
Series
Narrative build log spine: scaffolding, agents, docs, governance, and the daily operating loop.
Context infrastructure: Obsidian, Markdown, vault architecture, memory, and state.
Bounded agent roles, prompt contracts, handoffs, and safety boundaries.
Model evidence: where local models help and where frontier models still earn the electricity.
Reusable delivery patterns: scaffolding, reviews, docs, validation, and commits.
Lightweight controls: decisions, risks, permissions, and promotion criteria.
Enterprise architecture concepts at working size: capability maps, maturity, and dependencies.
The daily loop: backlog, briefings, context sync, and the rituals that keep one human and several agents aligned.
Promotion logic: when an experiment deserves productization, standardization, or retirement.