A graph that cannot say no
July 2026 filled the conversation with typed graphs for agents, on a thesis that is correct — an untyped edge carries one bit; a typed one carries meaning — and with independent benchmarks that back it. But those same benchmarks say something their popularisers do not finish: one system collapsed to 6.6 average F1 against 59.8 the moment somebody other than its authors evaluated it, and a five-hop chain at 85% per-hop accuracy is worth 44%. The problem is not missing structure. It is that structure, alone, obliges nothing. And the gap the article itself declares empty — a linter for typed edges — has been running here for a while.
The three chairs
Why would a project pause to declare its axioms, keep its tensions deliberately unresolved, and make every decision carry signed reasons — now, in a time of machines that write faster than anyone can read? Instead of arguing it, this piece stages it: an unnamed maintainer asks, Marcus Aurelius and Lao Tzu answer only with their own attested words, and the skeptics are not straw men in the prose but guests in the room. The Disruptor demands speed, the Legalist demands law, and Heraclitus — the master of the unity of opposites — attacks the serenity of the evening itself. Two objections get answered by the guests. The third is answered by the structure of the interview. The session adjourns with no verdict, tensions open, and one empty chair reserved for the reader.
The harness has its canon. Its open questions have answers.
The harness discipline now has a canonical text: guides steer the agent before it writes, sensors correct it after, and a steering loop improves both when issues repeat. This post adopts that vocabulary and takes the next step: the canon's own open questions — how do guides and sensors stay coherent as the harness grows, how do you evaluate harness coverage, what does a silent tool mean — are exactly the questions a typed, witnessed subject answers. A check that does not declare what it looked at is a harness lying by omission; a sensor whose verdict is signed is one you can audit. The harness runs the agent. The subject is what the harness queries.
Your subject is not for rent
The harness post left a question open: if your project's knowledge lives inside the harness profile, switching harnesses means losing it. The market's answer — move it out to an external knowledge base — solves one coupling by introducing another. This post walks the harness layer by layer through the pairings nobody decides and everybody assumes, and proposes the change of regime: from renting to owning. No villain: every step of the enclosure is reasonable on its own, and that is exactly the problem.
A harness without a subject: what harness engineering doesn't name yet
Harness engineering just named itself as a discipline: documentation as code, architectural constraints, layered verification, periodic consistency audits. All four pillars exist — in flat markdown, held up by manual discipline. This post concedes the whole argument (“the model is commodity, the harness is the advantage”) and adds the next step: the harness governs the verb — how a change is made — and names no subject to check it against. A verb without a subject cannot drift, because drift is precisely a subject diverging from its own declared self.
A rule without a trigger is a sign
No rules were missing. There were four — written, canonical, consultable — and not one of them ran. The distance between a rule that is written and a rule that applies has a name, and it is the only thing separating discipline from documentation.
Declared, Not Filled: Context You Can Audit
Jessica Talisman's essay splits the year's favourite buzzword in two: context as a container you fill, or context as a property you declare. ANSI/NISO Z39.19 has held the property view for a century — qualifiers, scope notes, warrant. ontoref reached the same record shape without ever citing the standard, then pushed it one step further: from declaring context on terms to declaring it on acts, where a signed witness answers 'on whose authority does this result exist?' with an exit code instead of fluency.
The Seven Sins of AI Agents
Agents don't fail at random: they fail with seven systematic vices that all survive the 'looks correct' test. The instinct is to add process — a PEP, a KEP, a committee — but every graduation stage rests on a human who approves it, and the agent's speed outruns the human you put at the gate. This is the honest comparison: what ontoref's ADRs inherit from PEP and KEP, and where they surpass both with witnessed, decidable, bounded-slice graduation criteria.
Trust Is an Output, Not an Input
An agent skipped the one invariant that would have caught the bug in thirty seconds. The honest diagnosis was not 'the agent forgot' — it was that the rule was prose, and prose never binds. This is the story of turning that failure into a falsifiable mechanism: a Statement of Work (the terms you own) and a Work Order (the execution it can't edit), where 'done' carries the validator's output instead of the agent's word.
Release engineering from one model
Release engineering is the most-copied, least-governed surface in any multi-repo org: CI YAML pasted between services until no two pipelines agree. ontoref models the entire release surface as one typed NCL workflow layer and generates the real artifacts from it — so the pipeline is queryable like any other architectural fact, and drift fails before it ships. ADR-038 added multi-arch OCI distribution as a catalog entry plus a generator, not a fork.
Your Ontology Should Live With Your Code
Enterprise knowledge graphs live in a triplestore maintained by a dedicated team on a vendor's platform. When the team changes, the budget gets cut, or the vendor pivots, the knowledge disappears. Ontoref takes the opposite approach: sovereign, local-first knowledge that lives alongside its subject, versioned, queryable across four surfaces, and federated without a central broker.
AI is a Knowledge Tool. But Who Keeps the Knowledge Alive?
Jessica Talisman's KGC 2026 talk is the clearest articulation of why AI strategies fail: organizations invest in data infrastructure and expect reasoning to emerge. She's right. But her solution — build knowledge infrastructure — stops exactly where the hard problem begins: who keeps the knowledge from drifting?