The arithmetic of autonomy is unforgiving. If each step in a task succeeds 95 per cent of the time, a twenty-step task completes correctly about 36 per cent of the time, because errors compound multiplicatively. This is why production agents are built from verifiable steps with checks between them, and why per-step reliability matters more than how impressive any single step looks in a demo.
Autonomy is granted by action class, not globally. Reversible, low-value actions — drafting, retrieving, summarising, updating internal state — run unattended. Irreversible or financially significant actions — payments, deletions, external communications — sit behind approval gates the agent cannot open itself. Iteration, tool-call, wall-clock and spend limits are enforced outside the model, because the characteristic failure of an autonomous system is not a wrong answer but an expensive loop.
Full autonomy works today where the loop is long and every step is checkable: research and synthesis, data processing, test-and-fix cycles against a test suite. It fails where goals are ambiguous and outcomes unverifiable, because the agent cannot tell done from plausible. The production requirements follow directly — complete traces, replayable runs, idempotent tools and a kill switch — none of which are visible in a demo and all of which decide whether the system survives contact with real users.
Deployments that succeed follow a trust ramp. The agent first runs in shadow mode — deciding but not acting, its choices scored against what humans actually did. It then acts under supervision, with a person approving each consequential step, and earns unattended operation action class by action class as the measured numbers justify it. Every expansion is reversible, and the evaluation suite arbitrates readiness rather than confidence or demo quality.
The human role does not disappear; it moves. People stop executing steps and start handling exceptions, reviewing approvals and supervising the system itself — which changes the product surface an autonomous agent needs. Less chat box, more operations console: queues of pending approvals, drill-down into any run's trace, cost and success dashboards, and clear ownership when the agent escalates. Teams that budget only for the model and tools discover this gap in production, where it is most expensive to close.
Codazz builds this in production — AI Agent Development.