The Agent Factory for Businesses

What We're Building

Ansur is platform business use to deploy company specific AI employees trained by their own team members, on their work. It takes the job function specification and generates the infrastructure required to make an AI agent trustworthy enough to run that job inside a real company. Not the agent itself, but the harness around the agent: evals, sandboxes, state machines, audit trails, supervision trees, and the human in the loop gates for irreversible actions.

You describe the job, and Ansur produces an AI employee with all nine of its components generated rather than hand built. The agent itself is disposable. The harness is the asset.

The Pipeline

The factory runs four steps. The first is ingest, where the system eats the company's SOPs, regulations, Slack history, and recorded interviews. The goal is saturation, meaning the system knows the job the way a new hire would after reading every document the company has.

The second step is decompose. The job gets broken into a workflow blueprint that separates the deterministic checks from the parts that need LLM reasoning gated by evals.

The third is generate. In a single pass, the system produces all nine AI employee components: agent logic, backpressure hooks, knowledge base, eval suite, audit trail, environment, persistent memory, supporting agents, and the human in the loop gates.

The fourth is the debug loop. The system runs the evals, diagnoses failures, refines the spec, and reruns, all autonomously. It iterates from first draft to production ready before a human looks at it. Once the agent is in production, corrections flow back into the eval suite. Fifty corrections in, fifty new test cases out. The harness gets stronger from real usage rather than from engineers guessing at edge cases.

Why It Holds Up

The architecture has three layers, each changing at a different rate. The bottom layer is infrastructure: the runtime, process isolation, audit trail, and persistent memory. It is stable and outlives any model generation. The middle layer is the meta harness itself, which adapts. When a new model lands, you update one place and every agent the system has produced gets regenerated against it. The top layer is the AI employees, which are disposable output meant to be thrown away when something better becomes possible.

Everyone else is hand crafting at the bottom layer and rebuilding from scratch every model release. We build at the middle layer, so model churn becomes a rerun rather than a rebuild.

The Bet

Agents are becoming a commodity. Every lab is racing to make agent deployment easier. The scarce resource is the infrastructure that makes agents trustworthy, and right now no one owns it.