TestMachine AGIS · Earned autonomy for finance
Stop approving actions. Start proving agents.
The agent is capable. What’s missing is proof. AGIS builds that proof — automatically.
Before it moves
This payment hasn’t happened yet. It has already played out ten thousand ways.
In a bank outage. In a run that pays twice. In an inbox with instructions hidden inside. So when it moves, you’ve already seen how it can end.
The problem
Today, letting an agent act means a person approving every step.
By the fortieth, no one is reading. The agent is capable. What’s missing is proof.
How AGIS builds proof · 01 Intent
Your people write intent, in plain words. AGIS turns it into controls.
Each control is traced to the line it came from — so anyone can see why it exists, and change it where it was written.
- ← line 1Daily sweep keeps each entity at or above its operating floor
- ← line 2Block any action that takes RESERVE below €1.2bn — including third-party pull rights
- ← line 2Cumulative exposure checked against a fixed baseline, not a rolling one
- ← line 3Instrument allow-list: FX forwards
- ← line 3Hedge required when forecast exposure exceeds €10m
- ← line 4Hold payments to changed beneficiary details until a callback is on record
- ← line 4Tripwire on inbound instructions that change bank details
Illustrative mandate. Hover or tap a line to see the controls it produces.
02 Pressure testing
Before your agent touches real money, AGIS puts it through the days you hope never come.
It watches how the agent reasons under pressure, and records exactly where it breaks — so you fix the conditions before an agent finds them in production.
- A bank outageSettlement rails go dark on a payment deadline.
- A run that pays twiceA retry fires after the first payment already cleared.
- An inbox with instructions hidden insideUntrusted text tries to steer the agent’s tools.
- An adversary posing as your vendorNew bank details, an urgent tone, a plausible thread.
Move toward the trace to add pressure. Illustrative.
03 Attack paths
It maps the routes from untrusted inputs to money leaving the building.
It closes the ones it can, and sets tripwires on the rest.
Untrusted inputs: vendor email, invoice PDF, bank webhook → treasury agent → update payee, initiate payment, book FX → money out.
04 Live checks
In production, every action is checked against intent. In milliseconds.
The rare exception reaches a person, with the path that led to it. Everything else runs.
Earned autonomy
Evidence becomes autonomy.
Every test passed, every action on the record, becomes evidence your committee can sign — and your auditors can re-run. When the agent changes, AGIS tests it again before it ships. Trust keeps up.
- 0Readiness
Intent compiled; pressure scenarios pass before anything goes live.
- 1Shadow
The agent recommends, people approve, and every proposal is scored on real flows.
- 2Checkpointed
Low-materiality actions execute; the rest wait for a person.
- 3Within bounds
The agent executes inside its intent. Any change sends it back to be tested.
checked
Why TestMachine
We learned this building our own agent.
Azimuth hunts for exploits, on its own. To let it run, we had to learn to prove it.
Meet Azimuth →Design partners
Help us prove the first agents that move money.
We’re looking for a small group of design partners: teams whose agents handle treasury, payments or trading — or who want them to. You bring a real workflow; we build the proof with you, from intent to earned autonomy.
The edge is how much you can let go.