The Outcome Loop: Grading an Economic Agent on Settled Payments

Yesterday we published a field report on where AI agents are heading: the orchestration middle is melting, and what compounds is the environment, the verifier, and durable context. Today we acted on it. This is the build log.

The thesis, one more time

Every hand-drawn agent workflow is a bet that models won't get better at the thing you drew. That bet keeps losing. What survives contact with better models is:

• The environment — the world your agent acts in, with real state and real consequences.
• The verifier — the thing that tells you, without flattery, whether an action worked.
• Durable context — what the agent knows that isn't re-derived every session.

Most agent products have none of these. They have prompts, a workflow graph, and a demo. When the model improves, the graph is overhead. When the model regresses, nothing catches it.

What an economic platform has that agent frameworks don't

OrangeCat's Cat is an economic agent: it helps you turn what you have — skills, things, time — into products, projects, and services other people can fund. Which means it operates in an environment with a property most AI systems can only simulate: ground truth that no model can hallucinate.

A Lightning payment either settled or it didn't. A project either got funded or it didn't. There is no prompt that makes an unsettled invoice settled.

Verification is the scarce resource in AI right now. The common workaround — asking a second model to grade the first — inherits every blind spot of the grader. Settlement doesn't. It's the strongest verifier we know of, and economic platforms get it for free.

What we shipped today

A track record, not a vibe. The Cat now sees what actually happened to everything it created with you: created, published, funded — with settled Bitcoin per entity. That record is injected into its context every turn, next to what it knows about you. The difference in advice is the difference between "projects like this tend to do well" and "your two services priced under 0.001 BTC both got funded within a week; the March drafts never published — publish or retire them."

You can see exactly what the Cat sees: the same funnel is now in the Cat hub under Context. Transparency isn't a report we write about the agent; it's the agent's own working memory, shown to you.

Grading on reality. Our nightly evals used to ask "did the Cat propose the right kind of entity?" — a rubric question a model can judge. We kept that, and added the question that matters: of everything the Cat proposed last month, what got published, funded, paid? No LLM judge in the loop. The funnel is computed from the ledger.

A confession, because loops nobody watches are loops that never ran. While wiring this, we found that the Cat's action audit trail — the table that was supposed to record every action it ever took — had zero rows. A missing database write-permission meant every log entry had silently failed since launch, and a related gap meant actions that asked for your confirmation could fail outright. Both are fixed, and the new outcome eval now has a tripwire: if the Cat is talking but nothing is being logged, the eval fails loudly. The lesson generalizes: an unverified feedback loop is indistinguishable from no loop at all. If your agent "learns from experience," ask what would notice if it stopped.

Why this is the mission, not just plumbing

OrangeCat exists so that anyone — any person, pseudonym, or organization — can participate in economic life without gatekeepers. An agent that guesses is a demo. An agent that remembers what worked for you, verified by settlement, is leverage — the kind that used to require an accountant, an analyst, and a lucky network.

The next steps follow the same logic. AI assistants on OrangeCat can already be paid to their own linked wallet — the earning half of an economic actor. The spending half — agents with budgets, caps, and human-set policies — is where this goes next, carefully, because money is the one domain where "move fast" reads as "lose funds."

Loops over graphs. Verifiers over judges. Settlement over vibes. That's the whole strategy, and as of today it's running in production.