Evidence for AI-agent decisions.
Agents in enterprise automation act on external facts. When one acts with real consequences, the question is never whether it was confident. It is what it knew.
One decision, end to end.
The decision
An agent approves an action
An autonomous agent is about to execute a consequential step: release a payment hold, place an order, escalate to a human. The approval consumes live external data at the moment it happens.
Evidence needed
What the approval relies on
Current market, weather, security or registry facts, each with a source, an observation time and a freshness state. The agent needs to know not just the value but how old it is and where it came from.
The failure
What stale evidence causes
An approval made on a value that was hours old, from a source that had already degraded, is indefensible in a postmortem. Without preserved evidence, nobody can show what the agent actually saw, so the argument becomes speculation.
Integration
How Dynamic Feed fits
The agent reads live signed data through the keyless MCP endpoint or REST. Every response carries provenance, freshness and an Ed25519 signature. Typical wiring is hours: one endpoint, no key management for the read tier.
The receipt
What remains afterward
A Decision Receipt: the exact inputs the agent saw, their sources and freshness states, the checks that ran, and the verdict, signed and timestamped. Anyone can re-verify it in a browser at /verify without trusting Dynamic Feed.
Pilot result
What thirty days should show
For each covered approval, a receipt exists and verifies independently, and honest freshness states are visible for every input, including the stale and unavailable moments. That is the measurable exit criterion.
Where to next.
The product tour walks one observation to an independently verified receipt · verify one yourself · the 30-day pilot · REST docs · MCP connection.