# Earned Autonomy

> A framework where agents earn the right to act alone, one proven action at a time, built on more than a decade of lessons in how people come to trust automated systems. Now open on GitHub.

- Organisation: Hivemind Labs
- Years: 2025 – now
- Role: Technical co-founder · framework author and researcher
- Try the demo: https://demo.hivemind.au
- GitHub repository: https://github.com/Hivemind-Labs-pty-ltd/earned-autonomy
- Read: Trust makes agents fast: https://jakespillane.com/writing/trust-makes-agents-fast

## At a glance

- **3×**: fewer serious mistakes than checks alone, across 6,000 tasks
- **73**: good actions in a row to prove an agent is at least 1-in-20 good
- **1 in 6**: errors that 20 passing runs can still hide

## The problem

Teams make AI agents safe the way they make any process safe: they add checks and reviews. Each one is a brake. An action that has passed the same check a thousand times still waits for it the thousand-and-first time, so the agent never gets faster.

The opposite mistake is just as common. Twenty good runs feels like proof, so the agent goes live. But twenty passes can still hide an agent that gets about one in six wrong. And a check only catches the mistakes someone thought to write it for.

## What I did

I designed Earned Autonomy around how we already trust people. A new starter gets checked closely; as they prove themselves, we check less. An agent works the same way, with two differences: its track record only counts for the kind of work it has actually been seen doing, and it needs more evidence than a person, because it rarely notices when it is out of its depth.

The business sets its appetite for risk per kind of task. Routine replies might tolerate one mistake in twenty; anything touching money, almost none. The agent speeds up only when its record statistically proves it is at least that good. Checks still run on everything, and anything that fails is held for a person. Trust decides how much of the rest a person still needs to see.

To test it, we ran real agents on 6,000 generated tasks and let the same share of their work through unchecked, with and without earned trust.

## The result

Agents relying on checks alone let through three times as many serious mistakes as agents that also had to earn trust. In a statistical simulation of the same rules, checks alone let through about twice as many.

The research is published, the framework is open on GitHub with a reference trust engine, and there is a live demo.

## What it taught me

Safety and speed aren’t a trade-off when trust is measured. The same evidence that keeps an agent safe is what lets it go faster.

Canonical page: https://jakespillane.com/work/earned-autonomy