An evidence check for agentic finance
An agent asks PolyBridge how likely the events behind a trading idea are. Then it previews a simulated trade, holds, or declines.
Published July 2026 · Read time 5 min
AI agents can now use tools. Give one a brokerage connection and it can move from an idea to an order very quickly.
That is useful, but it also creates a problem. The agent should not act just because an idea sounds reasonable. A trade is a claim about the future. Before the agent prepares an order, it should check that claim against outside evidence.
That is where PolyBridge fits.
Prediction markets put prices on future events. People have real money at stake, so those prices are a useful signal. But no single market prices most of the questions a trader actually asks.
PolyBridge fills that gap. It indexes more than 170,000 markets and models how the events behind them connect. An agent can ask any clear question and get back a probability, a range, and the source markets behind the answer, even when no market lists that exact question.
An agent can reason through a trading idea, but LLM reasoning alone is not an evidence check. Before the agent prepares an order, it needs an outside check: the question asked, the probability returned, the uncertainty range, and the markets behind the answer. PolyBridge supplies that record, so the decision can be audited later.
The Market Foresight cookbook shows this end to end. This article walks through a recorded run.
The thesis and its pass marks
Each question says which answer supports the idea and what probability would pass.
This is the thesis that ships with the demo:
{
"thesis_id": "labor-resilience-jul2026",
"as_of": "2026-07-04",
"thesis": "US labor market stays resilient through July 2026",
"instrument": "SPY",
"direction": "long",
"notional_usd": 1000,
"questions": [
{ "q": "Will the US lose jobs in July 2026?",
"supports_when": "NO", "threshold": 0.25 },
{ "q": "Will the US unemployment rate for July 2026 be above 4.3%?",
"supports_when": "NO", "threshold": 0.40 },
{ "q": "Will the Fed cut rates at its September 2026 meeting?",
"supports_when": "NO", "threshold": 0.30 }
]
}The idea is: if the US labor market stays strong, then job losses should be unlikely, unemployment above 4.3% should be unlikely, and markets should not be pricing an urgent Fed cut.
What the check does
PolyBridge Forecast answers each question with a probability, a range, and source markets.
If the evidence supports the thesis, the agent previews and places the simulated trade.
If the evidence clearly goes against the thesis, the agent writes a short memo and stops. It does not call the broker.
If the evidence is mixed, the agent holds.
Recorded run: the check passes
The probabilities below are recorded outputs from July 4, 2026.
| Question | Probability | Range | Pass mark | Result |
|---|---|---|---|---|
| US loses jobs in July 2026? | 12% | 8-15% | under 25% | Supports |
| Unemployment above 4.3% in July 2026? | 28% | 22-35% | under 40% | Supports |
| Fed cuts rates in September 2026? | 6% | 3-8% | under 30% | Supports |
All three answers came in under their marks.
The agent then previewed a $1,000 SPY buy and gave the reason. SimBroker recorded a simulated fill.
Conclusion
Agents are getting better at acting. The harder part is deciding when to act.
As agentic finance grows, agents will enter more financial workflows and higher-stakes decisions. The standard for what counts as enough evidence has to grow with it.
That standard is traceability: every probability used in the check should point back to real markets, and every decision should leave an audit trail.
Go further
• Cookbook: full pipeline, demo broker, and recorded fixtures.
Open the Market Foresight cookbook
• Run in Colab: open the notebook and run all cells. Open the Colab notebook
• Agent template: fill-in prompt for agents with their own broker tool. Open the agent template