product
One lifecycle around every agent run.
Verundo does not make agents smarter. It makes what they do plannable, controllable, observable and recoverable, across every system a run touches.
Understand risk before execution
The agent submits what it intends to do. Verundo judges the plan with the same rules that govern execution, so the answer before the run is the answer during it.
The plan comes back annotated: which steps will run on their own, which will wait for a person, which are refused, and how much financial exposure and customer-visible effect the run carries. Steps that cannot be undone are named as such before anything happens.
Teams use it two ways: agents that check their own plan and adjust, and reviewers who want to see intent next to outcome afterwards.
Policies, approvals and action controls
Every action executes through Verundo, under a policy the organisation owns.
- Policy by reversibility
- Actions are classified by what recovery is possible. Policy decides, per class and per action, what runs alone and what needs a human.
- Approvals in context
- A held step shows the approver the run so far, what the step would change, and what a rejection means for the steps after it.
- Anomaly detection
- Amounts and patterns outside the organisation’s own history are held, not silently executed.
- Attribution
- Every step is recorded with what it changed, and attributed to an agent identity and a model.
Stop runaway execution
A run can be contained in one click, by an agent that notices something wrong, or by a rule.
Containment is immediate and narrow: nothing further executes in that run, pending approvals are closed, and nothing is rolled back without a person deciding to. Rules can contain a run on repeated rejections, repeated anomalies, action rate, or cumulative financial exposure.
Reverse, compensate, surface
Recovery is honest about three kinds of actions.
- Reversible
- The before-state is restored and the restoration is checked against the live system.
- Compensatable
- There is no true inverse, but a compensating action exists: a cancelled subscription reactivated, an inventory adjustment offset.
- Irreversible
- A sent email, a refund the customer already has. Verundo never pretends. The effect is surfaced, counted and explained, and policy usually held it for approval in the first place.
- Order and dependencies
- Recovery runs in reverse and respects what depends on what, so a run is recovered as a whole rather than as a pile of unrelated undos.
Know when the world moved
Agents are not the only writers. A colleague edits the deal, a customer replies, another system syncs.
Verundo keeps watching the objects a run touched and knows when something else changed them afterwards. A recovery that would overwrite a person’s later work is shown to a person instead, with both states side by side, to keep or to revert. Where systems can push change notifications, Verundo hears about it in seconds.
Resolve uncertain outcomes
A request went out and no answer came back. That is not a failure, and it is not a success.
Verundo records the action as unknown, refuses to retry blindly, and reconciles with the provider to find out what actually happened. Only then does the action become executed or failed, and only a failed action is retried, without repeating the mutation.
Confirm recovery actually happened
A rollback that returned OK is not a rollback that worked.
After every recovery, Verundo re-reads the external system and reports verified only when the state matches. Each recoverable action also carries a recovery confidence, derived from what is observable: whether the before-state was captured, whether the object drifted, whether the system can be re-read and verified, how long ago the action ran. It informs; it never blocks a recovery.