A discipline for agents you leave running

Stewardship
Engineering

Operating an AI agent as a custodian of your work. Bounded authority on your behalf, accountable for what it does, keeping what you hand it over months of unattended operation.

Harness engineering
Gets the agent to act correctly while you watch. Build-time correctness.
Stewardship engineering
Lets you leave it running. Run-time accountability, earned over time.
The climb

The line that matters is the moment you stop watching.

Everyone using these tools climbs the same ladder. Rungs one to four teach the agent to act. Rungs five and six are about what it is allowed to do when nobody is looking.

Below the line, a mistake is an output you catch. Above it, a mistake is an action nobody caught. The loop gives you unattended execution. It says nothing about unattended authority.

  1. 01Prompt itPrompt eng.
  2. 02Feed it contextContext eng.
  3. 03Give it toolsHarness eng.
  4. 04Let it actHarness eng.
  5. You stop watching
  6. 05Run it unattendedStewardship
  7. 06Make it self-improvingStewardship
The keystone

Earned autonomy

Trust cannot come from how confident an agent sounds. So trust is made testable.

Before any real action, the agent writes a falsifiable prediction: a claim, the exact check, a baseline, a horizon. It goes to an append-only outcome ledger. Later a deliberately dumb verifier settles it, true or false.

Authority becomes a meter read per model and per task class. It is designed to widen as predictions hold and narrow the moment they slip. No certificate to bank. "Autonomous" and "unsupervised" are different words, and the distance between them is what gets earned.

Today the meter runs in shadow mode: it scores, it does not yet move permissions. The essay keeps the honest status.

Predict a falsifiable claim Act inside the envelope Verify a dumb check settles it Adjust the authority meter widen / narrow OUTCOME LEDGER · APPEND-ONLY every bet recorded before the action, settled after
The primitives

Six pieces, whoever's agent you run

Most are applications of patterns other people named. The outcome ledger is the new part.

Boundary

The envelope

A blast-radius boundary around what an unattended cycle may do. Reversible work runs free; irreversible work is staged for a human, until it is earned.

New

The outcome ledger

Agent-authored predictions, settled by a dumb verifier, feeding the trust meter.

Memory

Memory in custody

Knowledge durable across sessions and machines, not re-explained every morning.

Cost

Cost-tiered loop

A cheap deterministic gate every cycle; a capable model only when there is real work.

Routing

Model routing + council

Taint, then availability, capability, cost. A multi-provider council for load-bearing calls.

Recovery

State in git

The agent's whole working state as plain files, committed every cycle. Lose the machine, clone, relaunch.

Open source · MIT

The code

A discipline and a stack, not one binary, the way SRE is a set of practices and tools rather than a single repo.

Writing

The essay, and the work behind it

The essay Stewardship Engineering: How Agents Earn Autonomy Getting an AI agent to act is the easy half. Leaving one running for months, unattended and accountable, is the hard half. The discipline, the mechanism it turns on, and the system it runs on. Read on emmanuel.prouveze.fr →
Prior art

Standing on named ideas

In fragments, plenty of people got here first. What looks unclaimed is the synthesis, and the pre-action falsifiable prediction settled by an independent verifier.