The solution
What the factory builds
The solution is the platform I live in: my agents, the spaces they work in, the voice I call, the cards I answer. This page is the technical overview, for a reader who wants to know what it actually is. Concepts, not secrets.
It is one of two parts. The other is the factory, which builds and keeps this one.
-
A generic agent harness
An agent is a catalog entry plus instructions, skills, tools, delegation rules and visibility. That contract is the platform’s; nothing in it belongs to any one engine or model. My own agents are content on top of it, not edits to it.
-
Engines are interchangeable, and parity is a rule I hold myself to
Three engines run agents behind one contract: one on a vendor’s agent kit, one that speaks to many model providers, one driving a command-line tool. Moving an agent between them is a configuration change plus the known, documented gaps. Parity is a rule I hold myself to; known gaps are written down, and a capability that improves one engine while widening the gap to another is not finished.
-
The model is per agent
Which model an agent uses is its own configuration: a fast one for routine work, a stronger one where the stakes are high. No model or provider is written into the platform.
-
One door for AI outside the agents
Summaries, memory distillation, classifiers, reading images: every such call goes through one utility layer. Change the provider there and the whole platform moves with it, not just the agents.
-
Spaces, and departments of agents
Agents live in spaces. In a space a gatekeeper takes the request, generalists handle it, specialists are brought in, and a synthesizer writes the answer. They delegate to each other. Delegation across engines, so that which engine each one runs on does not matter to the others, is a target I build toward, not yet a given.
-
Configuration is an API before it is a screen
The platform is configured through a machine-readable surface first. A screen in the portal is one client of it, and optional. A core agent reads the same surface and guides a new environment through its own setup.
-
Storage behind an adapter
Files and documents sit behind a pluggable adapter: a network share, a document library, blob storage. The agents do not know which.
-
Every environment isolated by design
Each installation is its own set of resources with nothing shared between environments: not data, not secrets, not traffic. That isolation is a design rule: separate resources per environment, and a deploy aimed at the wrong environment is refused.
-
A person says yes
Risky actions stop at an approval gate. A decision arrives as a card with the agent’s recommendation, and a person closes it. Every action, by an agent or a person, is attributable to a person or to a named system principal.
-
True types, and gates in the build
A timestamp is stored as a timestamp, never as text that looks like one; a twenty-hour outage taught that. Rules that were repeated too often became build gates: no new tenant name or identifier in platform code, no new file past the size limit, no new text-parsed time, no new hidden timer that wakes the system. The gates block new violations; the old ones are listed, each with a ceiling it may not grow past.
-
A live evaluation suite
After each deploy, probes call the real chain the way the agents do, from the first door to the last store. Unit tests were green twice while the live chain was broken; that is why the probes exist. A red deploy reverts the merged group; a few live health probes can pause the queue; the other probes report and never pause; most of the suite is run on request.
None of this is for sale. It exists because I wanted to live in it, and because building it as if anyone could install it is how I keep it honest.