Agent
The agent that makes the call.
Metis Agent reads the world in real time, forecasts what's about to happen, weighs your options against your goals, and hands back the decision — then watches how it plays out and gets sharper.
- Grounded in forecasting models proven in the open
- A decision with its reasons — not a number to read
- Runs continuously and plugs into your stack
From data to decision, on repeat.
The agent runs the same loop every time it works a decision — and it compounds, because it remembers.
Reads the world
The live data you already watch, plus the signals you don't — weather, outages, markets, demand.
Sees what's next
Searches thousands of forecasts and scenarios, each with a calibrated confidence band.
Weighs your options
Scores every move against your goal and constraints, then tests them on data it has never seen.
Makes the call
Hands back a recommended decision with the reasons behind it — not a number to interpret.
Learns
Remembers what it recommended and what happened, so the next call lands sharper.
Built for decisions, not conversations.
The agent is engineered to be right when it counts — and to tell you when it isn't sure.
Grounded, not generated
The numbers come from forecasting models, not a language model improvising.
Calibrated confidence
Every forecast ships with a confidence band, so you know exactly how hard to lean on it.
Validated before it speaks
Candidate calls are back-tested on data the agent has never seen. It won't claim a win it didn't earn.
Searches the whole space
Thousands of forecasts and scenarios a run, hunting the move that actually wins — or telling you to wait.
Always on
Runs on your cadence — day-ahead, hourly, or on a trigger — and flags what changed since the last call.
Calls anywhere
Read it, let your team act on it, or pull it from your stack over API and MCP. Nothing to rip out.
Not another chatbot.
Chat assistants are fluent, but fluency isn't judgment — they can't tell when they're wrong, and high-stakes calls can't run on a guess. Metis Agent is grounded in the numbers, so the decision is one you can trust and trace.
A chat agent
- Sounds confident, but can't tell when it's wrong
- Made of words, not grounded in your numbers
- Answers a question and forgets it
- You still have to make the decision
Metis Agent
- Grounded in forecasts, baselines, and constraints
- Every call carries its confidence and its reasons
- Runs continuously and remembers every outcome
- Hands back the decision, traceable end to end
Proven in the open
We grade ourselves in public.
The forecasting under Metis Agent is benchmarked in public — on GIFT-Eval and live at tsfm.ai — and graded against the baselines buyers already trust. No black box. No numbers you can't check.
One agent, every desk.
The same loop, pointed at whatever decision you make under uncertainty.
Traders & desks
Position and hedge into the moves the market hasn't priced yet — sized to your risk.
Operators
Commit and dispatch against what the system will actually do, not yesterday's plan.
Researchers
Stop maintaining a model per series. Point the agent at the call and grade it against your baseline.
You point it at a decision.
No model to build, no pipeline to wire. Describe the call you make and the agent takes it from there.
Describe it in plain language
Tell the agent the decision you make. It scopes the target, the baseline you trust, and the cadence.
It runs on your schedule
Day-ahead, hourly, or on a trigger — the agent works the decision continuously and flags what changed.
Use the call anywhere
Read it, let your team act on it, or call it from your stack. Every decision is available over API and MCP.
Built on Metis-TSFM — see the models underneath.
FAQs
How is this different from ChatGPT or Claude?
Chat assistants generate fluent text and can't tell when they're wrong. Metis Agent grounds every call in forecasting models, baselines, and your constraints, validates it on data it has never seen, and hands back the decision with its reasons and a full trace.
Is it a black box?
No. Every recommendation carries its confidence, its drivers, and a trace from inputs to outcome — and the forecasting underneath is benchmarked in the open, not just on internal numbers.
Do I have to replace my models or systems?
No. It's decision support that plugs into what you already run — reports, API, MCP, CLI, scheduled feeds. Nothing to rip out.
How do you prove it beats what I have?
Two ways. Public benchmarks like GIFT-Eval and the live work at tsfm.ai, and a paid discovery against the baseline you already trust — leakage-safe, out of sample, with the worst day shown, not hidden.
What happens to my data?
You set the boundary and what the agent can use. Your data stays yours, and every recommendation traces back to the inputs that produced it.
How do I actually use it?
Through a workbench, a terminal and CLI, or your own stack over API and MCP. Describe a decision in plain language and the agent takes it from there.
Put the agent on a decision you make.
Bring one call you make today and the baseline you trust. We'll show what the agent does with it.