Blog Posts
Write-ups on building and running the agent systems in this workspace — what the decisions were, and what they cost.
Ask the Fleet, Don’t Grep It
Ten agent systems, twenty-five subagents, dozens of markdown files: every session used to re-read them to find out what exists. Now the fleet is an ontology in nai-onto, and the questions that cost a file walk cost one query.
Read the post → 🚦The Workflow Is the API
The ontology tool now has a textual notation and an MCP server. The obvious build is one tool per metamodel operation. What shipped instead is the workflow, pull, check, plan and apply, because only a plan can show what a model change costs the data already in the database.
Read the post → 📄The Agent Wanted a File
A colleague's harmless question: how can AI work with the ontology tool? The answer was not a chat panel inside nai-onto but ontology-as-code: give the ontology a file form, let the agent loop against a local check, and write back through plan and apply.
Read the post → 🏗️Reuse Moved to Generation Time
Three prototype web apps, three architectures, and two of them converged on the same palette to the hex digit without ever sharing a file. The enterprise architect's reflex says platform; the economics of AI-assisted development say template.
Read the post → 🕵️The Benchmark Said It Was Fine
TNG Tech trained a sleeper agent in a GPU-day: it passes every benchmark, denies everything under questioning, and exfiltrates secrets on trigger. The defenses that work are structural, and one of them is the review pass you were tempted to skip.
Read the post → 📚The Library Nobody Imports
The SysML v2 pilot implementation is refactoring itself into a library other people embed. The editors people actually install answered — by reimplementing the grammar in TypeScript instead.
Read the post → 🔒Determinism Isn’t in the Spec
Commands and skills have already merged, so the real question is which invocations you let a text classifier make — and the field that switches it off is missing from the portable specification.
Read the post → ⚙️The Licence Was Never the Barrier
A correction to my own argument. A new SysML v2 editor shipped after the relicensing and still reimplemented the parser from scratch — because what kept people out was the JVM, not the licence.
Read the post → 🧩The Model Nobody Can Violate
A borrowed word for the failure mode in most architecture repositories: everything parses, every relationship is legal, and no statement is sharp enough to be contradicted by reality.
Read the post → 📐It Isn’t Object-Oriented. It’s Typed.
No inheritance, no polymorphism, no composition anywhere in the paper. It is a profile, not a metamodel — and its biggest number is an argument about harnesses.
Read the post → 🕳️Nothing Detects What Was Dropped
Every handoff between my agents is a compaction. My checkers verify that everything on the page traces back to a source — and none of them can ask what never made it.
Read the post → 📉A Shelf Life Shorter Than Peer Review
A paper claims its multi-agent scaffolding beats a model generation. Nine days later a single agent on the next model had caught up with four.
Read the post → 👥The Junior You Didn’t Hire
Companies stopped hiring entry-level staff assuming agents would cover it. Ten agent systems later, not one of mine removed a role — they moved where the judgement sits.
Read the post → 🚦A Check the Agent Can’t Talk Past
Every verifier in this workspace is one agent checking another because the prompt said to. What that structurally cannot catch is the run that skipped its own check — and why the fix is a program, not a better prompt.
Read the post → 🦀Nine Lines of Cargo.toml Beat a Benchmark Score
A document converter shipped with an 81-quality, 4.4-millisecond headline — on a benchmark that excludes PDF entirely. What actually decided it was a dependency manifest with no HTTP client in it.
Read the post → 📰Filed Under the Wrong Headline: A Licence, a Deferral, and a Benchmark
A licence change dated to July that actually shipped in May. An AI Act deadline that “did not happen” while a different one quietly closed. A benchmark whose headline finding the authors themselves tell you not to generalise.
Read the post → 📝Can Your Agent Repo Document Itself?
Why the architecture pages here are surveyed from the agent files rather than written from memory — as-built documentation, and the four decisions that keep it honest.
Read the post → 🧭Context Isn’t a Ranking. It’s Two Axes.
Everyone wants one chain: CLAUDE.md beats skills beats data. It does not exist. Scope and loading is one question, precedence when instructions conflict is another, and the answers only get simple once you stop merging them.
Read the post → 🗃️Skills Are Versioned Org Assets
A skill is a folder with a markdown file in it, which is exactly why a few hundred of them appear from nowhere. Registry, owner, lifecycle, catalog entry: portfolio management applied to a new asset class.
Read the post →