Product
One control loop. Five parts.
Margin prices every agent operation against the outcome it produced: a merged PR, a passed test, a resolved ticket. Once a person approves it, the cheaper path is held only while it keeps quality. These are its five parts, in the order you meet them. Each one hands off to the next, and the last one re-arms the first.
- The Meter: measures every call, its cost, tokens, substrate, and whether it worked.
- The Estate: the org chart of your AI workforce.
- The Money Map: where the spend goes, decomposed by step.
- The Parity Gate: proves the cheaper route held quality before it ships.
- Auto-Tune: acts with a human ratifying, and a revert armed.
01 · You install
Every call, metered: cost, tokens, substrate, and whether it worked.
A small SDK wraps your LLM calls and reports the outcome when your own check already fires. It is fail-safe: a failed emit returns ok=False and never raises into your request path. It cannot spoof: the source is forced server-side to your project. Instrument one workflow first. The whole product waits for a definition of “worked”, because a guess in the denominator corrupts every number downstream.
02 · You see what you run
The org chart of your AI workforce.
Once calls are landing, you can see what actually runs: teams, environments, agents, operations. That is the shape of your AI workforce. Most teams cannot answer how many agents they have or what they are doing. The Estate is that answer, current to the last metered call: never a promise of “always live”, just the truth up to the last call.
See the estate →03 · You see where it goes
Spend is easy to total. Hard to locate.
The Money Map decomposes spend by step, so a rising bill has an address. The step names the operation, the task class, and the model. It surfaces the two numbers that matter first: cost per outcome, and the waste split between retries and cache. Both carry provenance: a metric is real or it does not render.
See the money map →04 · You prove it
A cheaper model that drops quality is debt, not a saving.
The Parity Gate runs one task class down both routes at once, the incumbent and a cheaper candidate, hundreds of times, and scores both against the frozen check you ratified in the scoping call. A route is promoted only when the cheaper option clears a high bar we hold ourselves to and costs less. If it does not hold, the verdict is defend: keep paying. That verdict is worth as much as the saving, because it is the first time anyone has proven a piece of your AI spend is earning its keep.
The parity engine, written up →05 · You act
The action half, with a human in the loop.
Once you ratify a promotion, Auto-Tune switches the route and arms an automatic revert. If parity slips on live traffic, it reverts on its own and says so. Autonomy is a dial you set, not a default we assume: the system proposes and proves, a person approves, and it undoes itself the moment quality drops. The ledger of every face-off compounds, so the tenth workflow is cheaper to evaluate than the first.
The closed loop, written up →Watch the loop run on real spend.
The console runs the whole thing live on open-source agents we metered ourselves: measure, recommend, act, revert, with the parity proof on every change and every number tagged real or sample.