AI · JUNE 14, 2026

Agents need permissions, not personalities

The interesting part of an autonomous agent is not what it can do. It is the exact list of things it is allowed to do, and what happens when it tries something else.

Sabbir Ahmed · 7 min read

More in AI

Agent frameworks sell autonomy. Businesses buy accountability. Bridging the two is mostly unglamorous engineering: scoped credentials, idempotent tools, retries with budgets, and an audit log that a non-technical operator can read.

Tools are an API surface, treat them like one

  • Every tool call is validated, typed and idempotent
  • Write actions are staged and confirmed before they touch a customer record
  • Budgets cap tokens, time and retries per run, runaway loops are a billing incident
  • Every run is replayable from its trace

Autonomy is a dial, not a switch

Ship at suggestion level first: the agent proposes, a human approves, and every approval becomes training signal about where it is trustworthy. Move each individual action to full autonomy only when the data says its approval rate is boring.