fyv
Evidence for what your agents say and do.
fyvAgents checks whether an agent’s claims are in the sources it cites, and keeps a record of what it did that the agent can’t change, so a person can always see why.
In development. Not live.
Built for
Risk, audit and compliance leads: check what any agent claimed and did, even one you didn’t build.
Engineers building agents: prove what your agent relied on and did, without asking anyone to trust its logs.
In
Federally regulated financial institutions: evidence of the kind a model risk review asks for, ahead of E-23.
Companies whose agents deal with other companies’ agents: who authorized it, what it relied on, and what it did.
Agent-to-agent business is growing
AI agents are software that act for a company: they answer, book, buy and negotiate. More and more, one company’s agent deals directly with another company’s agent. A buyer’s agent asks a supplier’s agent for a quote. A bank’s agent checks what a customer’s agent claims. When that works, it’s fast. When it goes wrong, each side needs to know who authorized its agent, what it was allowed to agree to, and what it actually did. Today, the only account of that is usually the agent’s own log.
Agents are starting to check other agents
Agents now hand work to other agents, and we expect more of them to learn from what they’ve already done. Agents like that change faster than the people responsible for them can check their work by hand, so oversight is moving to agents too: agents that check other agents’ claims, and agents that decide what other agents may do. We think that’s where this is heading, and it leaves one question open. Who checks the overseer?
An agent’s log is the agent’s own account of what it did, and an overseer’s log is no different. Every layer needs evidence it can’t change and a person can check. Oversight can be handed to agents. Accountability can’t.
Verify: is the claim in the source?
Your agent sends a claim and the sources it cited. Verify checks whether the sources support the claim and returns a signed receipt. When the sources don’t settle it, the answer is “cannot tell”. Verify doesn’t judge whether a source is true, only whether the claim is in it. Your agent runs it as a single check. It changes nothing, and it doesn’t go looking for sources on its own.
Record: a record the agent can’t rewrite
When something goes wrong, the first question is what the agent actually did. Record keeps a record of what an agent did, stored where the agent can’t alter it. It keeps a fingerprint of each action rather than its content, signs them into a chain, and is built so the record can be checked without fyv. An agent that oversees other agents is still an agent, so its actions go on the record the same way.
The same rules as our governed memory
fyvQuorum backs every claim with a citation and every action with an audit trail, and nothing enters a team’s memory until a person confirms it. fyvAgents takes those two commitments to agents we didn’t build: Verify for what an agent says, Record for what it does.
For a team, decision integrity means knowing what was decided, who may know it, and why. For an agent, it means knowing what it was allowed to do, what it relied on, and what it actually did.
We put the principle behind this to the federal government in September 2026, in our response to ISED’s consultation on AI agents: every agent should act under authority that a person or organization granted and recorded in advance, and an agent that hands work to another can’t hand over more authority than it has.
What it leaves to you
fyvAgents doesn’t decide what your agents may do. Your people and your policies decide that, and fyvAgents gives them evidence they can check.
How we hold ourselves to it
We don’t give an answer until we’ve tested it. We set the pass mark before we see any results, and Verify doesn’t go live until it passes. If a test fails, we say so.
Every receipt is signed and linked to the ones before it, so if a record is changed, it shows. fyv keeps the chain, and a reviewer can check it without taking our word for it.
It’s the same rule we hold everywhere: a claim you cannot audit is a liability that looks like an asset.
Who it’s for
Risk, audit and compliance leads at companies putting AI agents into work that matters, and the engineers building those agents. For federally regulated financial institutions preparing for OSFI’s Guideline E-23, it produces evidence of the kind a model risk review asks for.
If you want an agent built for your team, that’s fyvWorks. fyvAgents checks and records agents, whoever built them.
Building with agents? Help shape fyvAgents.