Agents that close the case.
Leri takes a request the whole way: reads it, finds the record, decides, writes the change back into your systems, and leaves the receipt. Eight faces, one loop, no handover unless it earns one.
Intake
A request arrives on whatever channel your customers already use. It is turned into a case with an owner, a clock and a state.
Read
Intent, entities and the thing actually being asked for. Ambiguity is recorded as ambiguity, not guessed away.
Retrieve
The record is pulled from your systems, not from the model's memory. Every field it reads is cited in the receipt.
Decide
Your policy, expressed as rules the agent cannot argue with. The decision names the clause it came from.
Act
The write happens here. Scoped credentials, one transaction, reversible, and nothing outside the scope it was granted.
Verify
The agent re-reads the record it just changed. If the world does not match the intent, the case does not close.
Log
One receipt per case: inputs, sources, clause, action, result. Readable by an auditor with no context.
Escalate
The eighth face is the honest one. When confidence or authority runs out, a person receives the case already worked.
One case, end to end, with the receipt attached.
Where it sits in your stack
Every action the agent can take is declared in advance, with limits, and nothing else is reachable.
Anything written can be replayed and undone from the receipt, without reconstructing what happened.
Resolution rate, escalation rate and time to close are reported per queue, not as one headline number.
The rest of the system
Own your own AI future.
Put an agent on one queue, with limits you set, and read the receipts.