The shift
Agents are moving from answers to action.
Capability is only part of the problem. Before an agent can do real work, it needs infrastructure beneath the model: a runtime that learns without gambling on production, company context with provenance and control, and a web it can understand without reverse-engineering pages built for eyes.
The constraint
Better models are not the same thing as reliable systems.
An agent can reason well and still use stale context, trust the wrong source, repeat an expensive path, exceed its authority, or misunderstand an action on a page. The failure is often not intelligence. It is missing infrastructure.
As soon as an agent acts, provenance, permissions, evaluation, fallback, audit, and recovery stop being backend details. They become part of the product a person or company must be able to trust.
The operating stack
Three parts of one operating thesis.
Punk, Pantasso, and Plasmate address three different boundaries beneath production agents; they are the stack I am building, not a claim that every agent needs only these three systems.
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
Autonomy should expand only as evidence accumulates.
Novel work cannot always be proven in advance. A trustworthy system should know the difference between what evidence supports, what remains uncertain, and what authority has actually been granted.
Knowledge with provenance
Context should identify its sources, respect permissions, and expose an honest gap when the evidence is not there.
Evidence before promotion
Repeated or cheaper execution paths should be evaluated before they replace live reasoning. Uncertain work should retain a safe fallback.
Scoped and recoverable action
Authority should be bounded. Consequential actions should be auditable, explainable, and recoverable when something fails.
The business consequence
Action changes the accountability.
When a vendor sells completed work or outcomes rather than seats, responsibility shifts toward that vendor—but this is a business-model choice, not the fate of every agent company.
Consequential work has to be reconstructable: what context was used, what authority was granted, which path was chosen, what happened, and how the system recovered when something failed. That is the technical foundation of the service-as-software bet.
Durable retention should come from accumulated context, evaluation, and customer-specific adaptation that can be inspected—not from making departure painful.
From thesis to operating stack
The writing formed the argument. Implementation is testing it—not proving the market.
The current season is concentrated implementation: turning the agent infrastructure thesis into durable, production-ready systems. The systems make different parts of the thesis concrete: how work is routed, what can be known, and how supported pages are made legible to agents.
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.
This is evidence of attention and implementation velocity. It is not evidence of customers, revenue, adoption, benchmark performance, or product quality.
Evidence has types. Implementation activity, retained benchmark observations, public specifications, and customer outcomes answer different questions. I do not treat one as evidence for another.
The intellectual archive
A reading path through the thesis.
The thesis did not arrive as a finished architecture. It developed through questions about runtime behavior, company context, the web, evaluation, governance, and service delivery. These are four useful starting points in an archive of 561 published essays.
- 01
A Season of Coding
How Punk, Pantasso, and Plasmate became one operating thesis.
- 02
The Runtime You Did Not Build
Why capable models still need an operating layer around them.
- 03
When AI Reads the Web Wrong
The case for web structure designed for machine understanding.
- 04
What an AI Agency Actually Needs
The evaluation, governance, and accountability behind delivered work.
The question
What has to exist beneath an agent before you can trust it to do real work?
Runtime. Context. Web. Autonomy that expands as evidence accumulates.
This is the thesis behind my current technical work.