01 How it works
Seven stages. One loop that never loses your work.
Every goal travels the same lifecycle. Here it is, walked through with a concrete example: a founder deciding between software vendors.
The goal, stated plainly.
Context gathered and kept.
Compiled, inspectable steps.
Routed to the right intelligence.
Every run receipted.
Checked against criteria.
Evidence becomes learning.
02 Walkthrough
“Which CRM should we buy?” — start to finish.
-
Work — the goal lands
You write: “Recommend a CRM for a 12-person sales team. Budget under $400/month. Must integrate with our billing.” That sentence, plus its constraints, becomes a first-class object — not a chat message that scrolls away.
-
State — context is assembled
The Context Compiler pulls what matters: your billing provider from company context, the team size from an earlier decision, the fact that you rejected two CRMs last quarter and why. Nothing you already decided gets re-decided.
-
Plan — Claude compiles the steps
The Work Compiler, running on the Claude API, produces an inspectable plan: gather requirements → shortlist candidates → score against criteria → draft recommendation. Each step lists its acceptance criteria, budget, and which steps need your approval. You read it before anything runs.
-
Execute — routed, not guessed
Reasoning steps go to Claude through the provider adapter; high-volume comparison grunt work drops to a cheaper tier. If a step needs a tool — a file lookup, a price check — it runs inside its permission boundary, and anything destructive pauses for your explicit approval.
-
Measure — the ledger writes receipts
Every model call is recorded: provider, model, tokens, latency, cost. The research run cost $1.12 across 14 calls — you see each one, not a lump sum.
-
Evaluate — checked, then judged
The draft recommendation is validated against your criteria: under $400/month? Billing integration confirmed? Then it lands in front of you for the only verdict that counts — accept, correct, or reject.
-
Recompile — the loop closes
Your decision, the costs, the corrections — all persisted as evidence. Next quarter, when the team asks “why this CRM?”, the answer is in the graph. And the Recompiler can compare this run against future ones on real numbers.
03 Posture
Modes set the posture. Gates hold the line.
The same lifecycle adapts: Quick for short answers, Deep when it has to be right, Research for evidence with provenance, Build for artifacts, Do for authorized actions, Private for restricted dispatch.
Approval gates
Publishing, spending, deleting, messaging — anything irreversible or financially consequential pauses for explicit human approval. Autonomy where it’s safe; a human where it matters. This isn’t a setting. It’s the architecture.
Run one real workflow.
That’s the alpha: one goal, end to end, with state you can inspect and receipts you can audit. Request access and bring a goal worth keeping.