How much autonomy should an AI agent have?

As much as it has earned for that specific task type, capped by how reversible the consequences are. Global autonomy settings are the wrong unit of control.

What sample size justifies more autonomy?

Around fifty reviewed instances of the same task type at 95%+ clean approval. Smaller samples produce confident promotions that reverse within a fortnight.

Should demotion be automatic?

Yes. Manual demotion never happens in time because nobody is watching the metric on the day it slips. Encode the threshold and let it act.

Is full autonomy ever appropriate?

For narrow, reversible, well-monitored tasks (scheduling, tagging, internal summaries), yes. For anything published under a person’s name, no.

Journal

AI agents

Autonomy Levels for Content Agents

Autonomy is not a setting you choose at install. It is a rating an agent earns per task type, and can lose.

Alex Cinovoj TechTide AIJune 16, 20263 min read

Key takeaways

Autonomy should be scoped per task type, never granted globally.
Promotion is earned by measured agreement rate over a real sample.
Demotion must be automatic, or the framework is decorative.
Reversibility and blast radius set the ceiling, regardless of accuracy.

Most products offer two settings: assist, and a toggle labelled something like "full autonomy" that nobody switches on twice. The interesting design space is entirely in between, and it is governed by two variables. How often the agent is right, and how bad it is when it is wrong.

The five levels

L0. Suggest. The agent proposes; a human does the work. Useful only as a research surface.
L1. Draft. The agent produces the artifact; a human reviews every one before anything happens. Default for anything new.
L2. Draft with silent approval. The agent produces and queues; a human has a window to intervene, and silence publishes. Cuts review load without removing the gate.
L3. Act with post-hoc review. The agent publishes; a human reviews a sample afterwards. Appropriate only where mistakes are cheaply reversible.
L4. Act. No routine review; monitoring and alerting only. Reserve for narrow, well-instrumented, low-consequence tasks.

The mistake is treating these as product tiers. They are per-task-type ratings. The same system might sit at L3 for scheduling, L2 for comment research, and L1 forever for anything that names a customer.

How an agent gets promoted

Promotion is a policy, not a conversation. For a given task type, track human agreement (approved without edits, approved with edits, rejected) over a rolling window.

promote(task_type) when:
 samples >= 50
 approved_without_edit_rate >= 0.95
 rejections_in_window == 0
 blast_radius(task_type) <= reversible

demote(task_type) when:
 approved_without_edit_rate < 0.90 (rolling 20)
 OR any rejection tagged wrong_claim

The demotion condition matters more than the promotion one. A framework that only goes up is a ratchet, and ratchets are how systems end up publishing something nobody would have approved.

What makes L2 the workhorse

L2 (queue, notify, publish unless stopped) is where most content operations should live, and it is under-used because it feels like a cop-out. It is not. It inverts the default from "nothing happens unless a human acts" to "the right thing happens unless a human objects," while preserving the human’s ability to object.

The design requirements are specific: a genuinely useful notification with the draft and its evidence inline, a window long enough to act in (we use four hours during working days), and one-tap stop. Get any of those wrong and L2 becomes L4 with extra anxiety.

The question is never "do you trust the AI." It is "what happens on the Tuesday it is wrong, and how fast do you find out."

Making it visible

Users should see the current level per task type, the agreement rate behind it, and the last promotion or demotion event with its cause. Autonomy that changes silently is indistinguishable from a bug, and the trust cost of an unexplained change is far higher than the trust gained by the automation.

It also makes the sales conversation honest. "It starts at L1 and earns its way up on your data" is a claim you can demonstrate. "Fully autonomous" is a claim that survives exactly until the first incident. Related: approval is the product and nobody wants another copilot.

The full autonomy ladder with the promotion criteria for each rung is at my site (opens in a new tab), and it governs how we ship agents at TechTide AI (opens in a new tab). Nobody starts at level four. Nobody should.

“Grant autonomy the way you grant it to a new hire: narrowly, per task, and revocably.”

Alex Cinovoj, TechTide AI

Questions people actually ask

How much autonomy should an AI agent have?: As much as it has earned for that specific task type, capped by how reversible the consequences are. Global autonomy settings are the wrong unit of control.
What sample size justifies more autonomy?: Around fifty reviewed instances of the same task type at 95%+ clean approval. Smaller samples produce confident promotions that reverse within a fortnight.
Should demotion be automatic?: Yes. Manual demotion never happens in time because nobody is watching the metric on the day it slips. Encode the threshold and let it act.
Is full autonomy ever appropriate?: For narrow, reversible, well-monitored tasks (scheduling, tagging, internal summaries), yes. For anything published under a person’s name, no.

Sources & further reading

Written by

Alex Cinovoj, Founder, TechTide AI

Alex Cinovoj builds ClawLI at TechTide AI, a fleet of specialist agents that research, draft, schedule, and grade LinkedIn content with a human approving every send. He writes about what breaks when you put agents in front of a real audience.

alexcinovoj.com techtideai.io See the pre-flight gate