Stewardship Engineering
A discipline and open-source stack for operating an AI agent as a custodian of your work: unattended, accountable, over months.
Operating an AI agent as a custodian of your work: running it unattended, accountably, over months. Bounded authority on your behalf, accountable for what it does, preserving what it is entrusted with over time.
The Problem
Getting an agent to act is the solved half. Mitchell Hashimoto named it harness engineering: building tools and environments so an agent acts correctly while you watch.
The unsolved half begins the moment you walk away. Putting an agent in a loop gives you unattended execution, not unattended authority. It will run a reckless action as faithfully as a careful one. What is the agent allowed to do at 3 AM? How do you know it is earning trust, rather than simply sounding fluent? And how does state survive when the host machine dies?
The Four Disciplines
The path from basic prompts to full stewardship climbs four disciplines across six rungs:
- Prompt Engineering: Instructing the model directly.
- Context Engineering: Curating and shaping what the model sees.
- Harness Engineering: Giving the agent tools, tests, and runtime environments so it operates correctly while supervised.
- Stewardship Engineering: Giving the agent bounded, revocable authority to operate unattended, test its own predictions against operational reality, and preserve institutional knowledge across machines and sessions.
The critical threshold sits between harness and stewardship: the boundary where you stop watching. Below that line, an error is caught in real-time. Above that line, an error is an action that ran without supervision.
The Keystone: Earned Autonomy
Trust cannot be granted once as a blanket permission, nor can it be based on model confidence. Models are fluent rationalizers of their own errors.
Stewardship Engineering turns trust into an empirical measurement:
- Pre-action prediction: Before executing a significant action, the agent writes a falsifiable prediction to an append-only outcome ledger (claim, check command, baseline, horizon).
- Independent settlement: A deliberately simple verifier executes the check after the horizon and records whether the bet passed or failed.
- Dynamic authority meter: Autonomy is computed per model and per task class. The blast radius widens as predictions hold, and contracts the instant they fail.
The Stack Components
Stewardship Engineering is an architectural doctrine implemented through three core open-source pillars:
| Component | Repository | Role |
|---|---|---|
| Heartbeat | eprouveze/heartbeat | The autonomous execution loop. Enforces hard blast-radius envelopes and run gates. |
| RightModel | eprouveze/rightmodel | Routing cascade, standing multi-model council, and the falsifiable outcome ledger. |
| Anamnesis | eprouveze/anamnesis | Distributed memory substrate with zero single-point-of-failure. |
Open Source & Reading
- Umbrella Repository: github.com/eprouveze/stewardship-engineering (MIT)
- Manifesto & Architecture Essay: Stewardship Engineering: How Agents Earn Autonomy