The shift
Agents are moving from answers to action.
Capability is only the beginning. The moment an agent leaves the chat box and starts doing real work, it needs more than a better model. It needs a runtime that can learn without turning production into a bet, company context it can trust, and a web it can understand without first digging through everything built for our eyes.
The constraint
A smart agent can still make a mess.
An agent can reason its way through a hard problem and still use the wrong document, trust an old answer, repeat an expensive path, act beyond its authority, or click the wrong thing on a page. The failure is not always intelligence. Often, the missing piece is everything around it.
As soon as an agent can act, provenance, permissions, evaluation, fallback, audit, and recovery stop being backend chores. They become the difference between a demo and a system someone can trust.
The operating stack
Three systems. One question.
Punk, Pantasso, and Plasmate look like three products because they work at three different layers. I am building them as one stack, aimed at the same question: what has to exist beneath an agent before I would trust it with real work?
adaptive runtime
Punk
Punk observes expensive repetition, proves cheaper execution paths against evidence, governs risky actions, and preserves the live model as fallback.
governed company context
Pantasso
Pantasso turns approved chats, documents, files, and business systems into cited answers, canonical records, honest gaps, and scoped access.
agent-readable web
Plasmate
Plasmate compiles pages into a Semantic Object Model so agents can work with structure, intent, and affordances instead of raw HTML and pixels.
The operating discipline
Trust should be earned, not switched on.
Some work is new. You cannot prove every path before the agent takes it. But the system should know what it has seen before, what the evidence supports, where the uncertainty begins, and how much authority it actually has.
Show your sources
An answer should point back to where it came from, respect who is allowed to see it, and say so when the evidence is not there.
Prove the shortcut
Repeated work can become faster and cheaper, but only after the new path earns its place. Until then, keep the live model available.
Keep a way back
Risky actions need hard boundaries. When something goes wrong, the system should be able to explain what happened and recover from it.
The business consequence
Sell the outcome and the accountability moves with it.
When a vendor sells completed work instead of software seats, the customer is no longer paying for access. They are paying for the result. That shifts responsibility toward the vendor. It is a business-model choice, not the fate of every agent company.
If the work matters, you have to be able to reconstruct it. What did the agent know? What was it allowed to do? Which path did it choose? What happened next? That is the technical foundation beneath the service-as-software bet.
Retention should come from a system that learns the customer and gets better at the work. It should not come from making departure painful.
From thesis to operating stack
Writing made the map. Building is finding the weak spots.
The current season is concentrated implementation: turning the agent infrastructure thesis into durable, production-ready systems. Each system forces a different part of the argument to become specific: how work is routed, what the agent can actually know, and how it understands the web outside the company.
Build signal · updated August 2026
Punk and Pantasso remain in sustained production hardening; July became the busiest recorded private contribution month at 3,927, moving the private rank from #1,769 on July 17 to #1,285 on August 1.
Commits show where attention is going. They do not tell you whether customers care, whether the products work well, or whether the market is there.
A benchmark, a commit graph, and a customer result are different kinds of evidence. I keep them separate because collapsing them is how honest work turns into marketing.
The intellectual archive
The thesis arrived one question at a time.
I did not sit down and draw the finished architecture. I kept running into the same questions about runtime behavior, company context, the web, evaluation, governance, and delivered work. These four essays are a useful way into an archive of 562 published posts.
- 01
A Season of Coding
How three systems became one operating stack.
- 02
The Runtime You Did Not Build
Why shipping an agent and running one are different problems.
- 03
When AI Reads the Web Wrong
Why the format of the web keeps making agents work too hard.
- 04
What an AI Agency Actually Needs
What changes when the customer pays for the result.
The question
What has to exist beneath an agent before you can trust it to do real work?
Runtime. Context. Web. Then more autonomy as the evidence earns it.
This is what I am building now.