A platform for building, running and supervising multi-agent AI workflows in regulated enterprises, where "the AI handled it" is not an answer you can put in an audit.
Every agent tool ships the same thing: a chat box and a promise. Fine for a demo. It falls apart the moment a regulated team asks the only question that matters. What happened, and can I prove it?
Agents delegate to agents. Steps fail halfway. The same input can take two different paths on two different days. For a compliance team, "the AI handled it" is not something they can hand to an auditor.
I made the workflow the object, not the conversation. Every run is a graph you can read. Each step carries a state, a confidence, an owner, and a defined fallback for when it breaks.
Failure is a first-class state, not an error toast. Humans approve at gates the team sets, and every approval is logged. The hard part was density without noise: a live workflow throws a lot of state at you, so the job was making consequences obvious at a glance and letting people drill in only when a decision is needed.
Four surfaces a risk team can operate: the orchestration canvas, a live run with real failure recovery, the empty state that gets a first workflow live, and the governance layer where guardrails and approval gates live.
On a fictional pilot, review time on an automated claims workflow dropped from days to under an hour, with a complete audit trail per run.
policy_api timed out twice. Per this workflow's rule, a failed policy check can't auto-proceed. The run is holding and Fraud signals continues in parallel. Choose how to recover:
Start from a template built for regulated ops, or open a blank canvas. Either way, guardrails and audit are on by default.
Where a human has to sign off before an agent workflow can continue. Rules apply across every run in this workspace.
| Time | Actor | Event | Run |
|---|---|---|---|
| 14:22:07 | E. Muller | approved payout · €4,120 | r_8f2a |
| 14:21:40 | agent · policy_check | failed → routed to human | r_8f2a |
| 14:19:02 | system | PII redacted (3 fields) | r_8f2a |
| 13:58:11 | J. Kaur | edited guardrail · threshold €5k | — |
A Very Vinci concept study. Invented brand, real design thinking.