Ideas from the team building infrastructure for governed agentic systems.
A mark is worth what it costs to withhold. Let the judge sell nothing he judges, keep his refusals where they can be read, and answer for a false mark with his own ruin. The sixth and final fable from Fables of the Walled Garden, with the collection’s afterword.
Perfect execution of a stale intent is a slow-motion failure that passes every review. Someone must own the distance between what was asked and what is now needed. The fifth fable from Fables of the Walled Garden.
“Whether the website allows it” is three different questions with three different kinds of consequence: the seller’s terms, the seller’s security system, and the law. The terms, the pause, the BOTS Act, and the Ninth Circuit opinion that didn’t move the responsibility.
Any contestant may excel on a course he drew himself. Before trusting a chart, ask who chose the axes — and whether the judge can reach the grapes he is scoring. The fourth fable from Fables of the Walled Garden.
The refinery started selling cars. That changes who owns the engine. On Matteo Wong’s Atlantic piece, the September incidents, and the engineering standard the accountability argument needs.
A signal can arrive perfectly while everything it was meant to signify has already been eaten. When the word “compliant” costs nothing to say, it says nothing — the question is what the crier must witness, and what happens to him when he cries it falsely. The third fable from Fables of the Walled Garden.
Every general-purpose technology has had the same three acts. The third one doesn't fix the machine. It makes the machine answer to a system it doesn't control.
The FDE is not a new go-to-market motion. It is the oldest one in commercial computing, and it appears when a capability has outrun the market’s ability to contain it. IBM invented the job in 1964. The industry just forgot.
An inspection the inspected cannot fail is not an inspection. Count the failures in the ledger; where there are none, there has been no judging — only signing. The second fable from Fables of the Walled Garden.
He who cannot move his beanstalk does not own it. He who cannot carry his evidence out of the garden never had evidence — he had a receipt. The first fable from Fables of the Walled Garden.
What the human meant, what the model received, what the system did, and what can be proven are different things. A governed system keeps them connected. The formal framework behind Equilateral’s architecture: Authority, Controls, and Evidence across the H→D→P→S→A→R pipeline.
Shopify’s CEO is right about the complexity tax. The filename is only where it surfaces. The real problem is an architecture that chose drift as its default state and then hired a robot to sweep up after it.
Every component in the MCP ecosystem can be locally correct while the resulting system is globally unauthorized. Authority that isn’t preserved end-to-end isn’t authority — it’s a crossed gate.
The structural case for external AI governance. The model cannot be a reliable governance authority over itself — not because of capability limits, but because of architectural ones. Three failure modes, five architectural primitives, and the admissibility thesis: capability gets the work done, evidence gets it admitted.
Four signals in one week. Illinois mandated independent third-party AI audits. GOLD EAGLE stood up a federal evidence clearinghouse. AAIF is standardizing agentic decision formats. The UK published its defense plan. The certification market just became mandatory.
Explainable. Auditable. Accountable. Every AI governance framework uses these words. The problem is they mean different things to different people — and the common interpretations fall short of what enterprise deployment actually requires.
In February I said we were promoting the LLM agent from tool to container before the governance layer existed. What I didn’t know: the transition was already complete. We’re inside the ungoverned container now. The window between promotion and liability is compressing from years to quarters.
The industry is optimizing the wrong layer. Model governance asks whether a model is safe to run. Decision governance asks whether a specific decision was authorized — under what constraints, with what evidence. Almost the entire field is focused on the first question.
Intuit compressed months of OBBB tax code implementation into days using Claude plus custom infrastructure. They built a bespoke knowledge engine by hand. Here’s the productized version — and the model drift problem they didn’t mention.
The summarizer becomes the sender becomes the transactor. AI agents escalate from T1 reading to T4 autonomous action one feature at a time — but governance is reviewed at launch, not at each capability expansion. The gap is your risk.
Anthropic’s 1M token context window makes the problem worse, not better. Your governance rules are tokens competing for attention weight — and they’re losing. The solution isn’t a better text file. It’s an architecture.
Anthropic’s RSP v3 is an honest post-mortem. The model maker couldn’t govern the model through the model. The lesson for every enterprise deploying AI at scale: build the external governance layer.
A system that cannot forget cannot govern. Enterprise AI governance requires both provenance and decay — capture with attribution, curate with evidence, forget with evidence. The WSJ described half the problem. This is the other half.
The Wall Street Journal describes enterprise AI capturing employee knowledge without governance. The answer isn’t to stop capturing — it’s to capture correctly. Structural attribution, earned maturity, governance by constraint.
The open-source coding agent ecosystem solved execution. It did not solve institutional memory. MindMeld is the missing layer between agents and governed enterprise deployment.
The Anthropic-Pentagon controversy exposed a structural pattern: governance by policy drifts under pressure. Governance by architecture holds. Here's why every enterprise AI buyer should know the difference.
CircleCI's data across 29 million CI workflows shows main branch success rates at a five-year low. Nearly one in three merges fails. The model isn't the variable. The architecture is.
Six hundred years after mastering the explosion, someone built the engine. We are in the fireworks phase of agentic AI. Everyone is impressed by the explosion. Nobody has built the chamber yet.
Every generation promotes impressive technology from tool to container. OLE, Flash, RPA, and now LLM agents. The pattern is identical. The resolution is older than the problem.
There is a 1968 animated film that most people remember as a psychedelic curiosity. They missed the architecture. An allegory for why governed AI requires a submarine, not a fleet.
The AWS Kiro incident exposed the gap between build-time configuration and runtime authority. Agent governance lives in three layers. Most platforms only have one.
AI agents are contingent workers. The governance expectations should match. We built an open scorecard—6 dimensions, 20 criteria—to make that evaluation concrete.
Cursor's research showed agents spiral in endless correction loops. We built an open-source solution: inject standards before the first token is generated.
AI systems are becoming unbiased record keepers. Whether that exposes the humans behind the system or the humans using it depends entirely on how we build them.
34 researchers from Stanford, Harvard, Berkeley, and Caltech explain the adaptation gap. We mapped our production system against their framework.
The scaling era is ending. The architecture era is beginning. Why frontier models need governance infrastructure, not bigger parameters.
At re:Invent, Werner didn't hype AI. He described the control plane problem that most agent architectures ignore—and why governed autonomy is the path forward.
No single model should have unilateral authority over critical decisions. Here's the architecture that prevents single points of failure.
Occasional updates on governance, architecture, and the future of autonomous systems.
Subscribe