Why we built VeriProof
Agents work in the demo and then stop moving. Some of the ones that ship get pulled back later. What sits behind both, and what we built about it.
Field notes
What it takes to run agents in production
Notes on getting agents approved, keeping a growing fleet under control, and answering for what they did. Each one starts from a question someone in the room has to answer.
Start here
What happens between the agent and the people who answer for it
One claim payment the agent could not decide alone, followed through the six roles that each had a part in it.
Getting to productionPolicy and decisionsPeople and handoffsEvidence and auditAgents across the businessRegulation
Getting to production
Most agent programs have a working pilot and no rollout. These are about what stands between the two.
Why agent pilots stall
The model is rarely the reason a rollout stops. Work stalls when nobody can say which rule applied, who approved the exception, and what the agent did.
The real cost is the second application
If every new agent costs what the first one did, the program stops at however many projects finance will approve.
Policy and decisions
What a rule should return when an agent proposes something, and what the application does with the answer.
Allow and deny are not enough
A discount over guideline is not a security incident. A large payment is not an error. Between yes and no there are three answers most systems cannot express.
Rules that read like your industry
When every team invents its own field names, the agent sends one thing, the policy checks another, and the reviewer reads a third.
People and handoffs
What happens when a decision has to reach a person, and how the answer gets back to the work that was waiting.
Escalation without authority is a ticket
Routing an exception to a person is easy. Routing it to someone who can decide, and getting that decision back into the workflow, is the hard part.
What happens between the agent and the people who answer for it
A developer, a governance engineer, a reviewer, a compliance officer, a business owner, and an auditor, each doing their part in their own portal.
Evidence and audit
What an auditor accepts, and why the thing your system already writes is usually not it.
A log is not evidence
A trace is an account you write about yourself. Evidence is something a stranger can check without trusting the party that produced it.
How VeriProof checks evidence integrity
The signed export, the Azure SQL Database Ledger check, and what a reviewer can confirm without calling us.
Agents across the business
Controls that reach one framework leave the agents built elsewhere running without them.
The agents you did not write
A colleague builds one in a maker studio. Another works through a browser. A control that only wraps your own code never sees either of them.
Rules for the framework your team already uses
The 1.0.0 release puts the check inside LangGraph, CrewAI, Semantic Kernel, Spring AI, and the other frameworks your developers already chose.
Regulation
What specific obligations ask for, in terms an engineer can build against.
What Article 12 asks for
"Automatically generated logs" is four words carrying a lot of weight. Here is what it means for a system where an agent proposes and a person approves.
Put one of these to the test
See a week of one agent's work and the record behind each exception, or evaluate your own agent with Developer, free for 60 days.