How much autonomy does a workflow really need?

How much autonomy does a workflow really need?

Autonomy is not a feature to maximise. It is delegated authority: permission for a system to interpret a goal, choose steps and change the world without asking first. The right level therefore depends less on model intelligence than on the consequences of a wrong action.

Begin with the workflow, not the agent. Stable rules, known inputs and limited exceptions still favour deterministic automation. Reasoning adds value where the route cannot be fully specified in advance: investigating an anomaly, reconciling conflicting evidence or adapting a plan. Most useful designs are hybrid�fixed controls around a narrow zone of machine discretion.

Set an autonomy budget from three variables: impact, reversibility and detectability. A low-value action that is immediately visible and easily undone can run unattended. A payment, customer promise, access change or regulated decision should face tighter limits, even when accuracy appears high. Human oversight belongs at consequential branch points, not after every harmless step.

Real-world behaviour is already changing. A 2026 analysis of millions of agent interactions found that the longest coding sessions grew from under 25 to over 45 minutes in three months; experienced users approved more work automatically, yet also interrupted more often. Oversight is shifting from approving every action to observing trajectories and intervening quickly.

That requires an architecture of constrained agency: least-privilege tools, transaction and data boundaries, explicit stop conditions, durable logs, checkpoints, idempotent actions and a tested kill switch. Measure straight-through completion together with exception rate, human corrections, rollback cost and time to safe recovery. Expand authority only when evidence shows that the whole workflow�not a polished demo�remains controlled.