Why we built agents around approvals, not autonomy
Fully autonomous agents make great launch videos and bad colleagues. Our Enterprise Agents are designed around a different assumption: a human should be able to stop, approve and understand every consequential action.
When we automated manual work for companies one workflow at a time, the pattern was always the same. The automation handled 80–90 percent of cases cleanly. The remaining cases — a new supplier, an unusual amount, a customer who wrote in a language we hadn't seen — were where the value and the risk both lived.
Enterprise Agents are built for that shape of work.
Read freely, write carefully
Every tool an agent can use has read and write scopes, granted separately. Reading a mailbox or a CRM is low-risk and reversible; posting an invoice or sending an email is not. We default to read scopes and ask administrators to grant write scopes deliberately.
Approval policies are rules, not vibes
An approval policy is a condition on a tool call: amount > 5000, supplier.new, recipient outside company domain. When a call matches, the agent pauses, packages what it found and why it wants to act, and asks a named group in Slack or Teams. One tap approves. The run continues with the approval recorded in the trace.
Policies live in the agent's configuration, versioned with everything else. Nobody approves a run by editing a prompt.
Exceptions are a hand-off, not a failure
When the agent can't complete a case within its confidence and its rules, it doesn't guess. It writes a case summary — what it read, what it tried, what's missing — and hands it to a person or a queue. In practice, this is where teams see the biggest change: staff review a handful of exceptions instead of processing every item.
Evaluate before you switch it on
Every agent can be run against a dataset of past cases with known outcomes. We compare the agent's decisions with what humans did and report accuracy per case type. Teams set a minimum accuracy gate; an agent that fails can't be promoted to production.
After release, sampled runs keep feeding the same evaluation so accuracy drift shows up before customers notice.
Limits, because software has bugs
Volume limits, spend limits and allowed hours are set per agent. An emergency stop halts any agent — or every agent in a workspace — immediately. We hope you never need it; it is there because we have been on the other side of a runaway automation.
Autonomy is a spectrum, and the right point on it depends on the workflow and the consequences. Our job is to give operations teams the controls to move along it deliberately.