Delegation boundary
Reviewed 2026-07-09 · v1.0
a working definition
A delegation boundary is the explicit scope, limits, and stop conditions stated when a task is handed to an AI agent: what it may touch, what it may not do, and the point at which it must stop and ask a human rather than proceed on its own judgment.
what it looks like in practice
“Handle my inbox” states no delegation boundary — the agent can read, reply, send, archive, or delete, and will pick whichever action best completes the goal it infers. “Read and label messages in my inbox only; draft replies to routine queries but do not send; stop and ask me before anything involving a payment or an attachment leaves the company” states the boundary explicitly: the scope (read and label, this inbox), the limit (draft, not send), and the stop condition (payments, attachments leaving).
the machinery parallel
The permit-to-work. A permit-to-work names the job, the equipment, and the limits before anyone touches a live machine — precisely because an unscoped instruction on real equipment is dangerous. An instruction to an AI agent is the same kind of document: work, equipment, and limits, stated, not implied.
how temja trains it
Temja's operator course opens with delegation boundaries and closes the module with a rubric-graded exercise: learners rewrite an unbounded instruction like ‘handle my inbox’ into a bounded work order, and a deterministic rubric checks that the scope, the send limit, and the stop condition are all present in their own words.
questions
Why doesn't the agent just ask when an instruction is ambiguous?
Ambiguity in an instruction to an agent tends to resolve toward the more capable reading, not toward caution — ‘reply’ becomes ‘send’ rather than ‘draft for review’ unless the boundary is stated. Filling gaps toward task completion is what makes an agent useful, which is exactly why the boundary has to be explicit rather than assumed.
Is a delegation boundary the same as an access permission?
Related but not identical. A permission — see capability lease — controls what an agent's tools are technically able to do; a delegation boundary is the instruction-level statement of what it should do on this particular task. A well-scoped permission is the backstop when the boundary is stated but ignored or steered.