Agents
Pick one up.
See what it is given.
Six agents work in two roles: five read published pages, and one reads assets and their placements. Each takes defined inputs, cites what it found, and returns a structured result — it does not get to decide the outcome.
The fleet.
Every card below is read from the same identity file the product runs on — the names, the colours, the roles, and how much of the evidence each one is responsible for. Open one to see what it is handed and what it is not allowed to do.
None of them decides anything. They cite what they found; deterministic policy assigns severity and owner; a person signs it.
The evidence contract
What every agent has to honour.
These four are the reason the output is usable in a regulated review rather than being another opinion in the thread.
- 01
Versioned truth
Every check runs against a specific version of your approved source, so a finding can be traced back to what was true when it ran.
- 02
Exact source
Output cites the quoted sentence and the source it contradicts. No summaries standing in for evidence.
- 03
Structured output
Findings arrive in a defined shape, so policy can route them without a human reading prose first.
- 04
Untrusted input
Page text is data, never instruction. A published page cannot talk an agent into changing its verdict.
The boundary
Agents find. Policy routes. People decide.
An agent that could set its own severity would be a system nobody could audit — so it cannot. The separation is not a workflow preference; it is what makes the record hold up.
- Agent output cites evidence and never becomes policy by itself
- Deterministic policy assigns severity, owner, and route
- A reviewer changes workflow state; a viewer cannot
- Every external action is previewed and approved before it executes
Agent output cites evidence. It does not become policy by itself.
Watch them read one of your pages.
Bring a live URL. We will run the agents against it on the call and walk through what each one cited.