On TheAgentCompany's workplace benchmark, the best models score 70.8% on a bounded task like requesting time off but drop to 35.3% on complex customer routing. Same models, same environment. What changed is the work surrounding them.
Coding agents scaled fastest for a version of this reason. Repositories already supply the accountability rails: tests, diffs, CI, review, reversibility. The agent enters a system that can define completion, verify outcomes, and undo mistakes. That's a property of the workflow.
So what determines whether work is delegable has less to do with model capability than with how crisply the surrounding system defines success. Work that has a stable case identity, verifiable completion, scoped authority, and a recovery path is delegable now. Work where exceptions live in someone's head and final state requires human inference isn't, regardless of how powerful the model gets. The more useful question for organizations is whether the surrounding system can tell if the agent did it right.
Is this workflow ready to delegate?
| Signal | Delegable | Not yet |
|---|---|---|
| Case identity | Stable ID in a system | Lives in email threads |
| Completion | Verifiable final state | Requires human inference |
| Authority | Scoped credentials | Shared or implicit |
| Exceptions | Documented paths | One person "just knows" |
| Recovery | Rollback defined | Manual or undefined |
Why it matters: An agent that succeeds 61% of the time on a single attempt may fail 75% of the time across repeated production runs. System-level constraints are what close that gap.