A lightweight pattern and reference implementation for governed collaboration between agents and the systems they control. It packages a clear governance loop—Propose → Explain → Confirm → Execute → Learn—so teams can ship automation without surrendering agency.
Alpha — The core loop is implemented and tested. APIs may change. Feedback welcome.
Automation systems and AI agents increasingly propose or execute actions with real-world consequences. This creates operational risk when:
- An agent proposes a destructive action without human review
- Audit trails are incomplete or missing
- There's no clear rollback path for failed actions
- Rate limits or blast radius controls don't exist
The stewardship layer addresses these by inserting a gate between intent and execution. Every action flows through Propose → Explain → Confirm → Execute → Learn. A steward (human or authorized agent) retains final authority. All decisions are logged.
Concrete use cases:
- CI/CD pipelines that require approval before deploying to production
- Home automation (Home Assistant, IoT) with confirmation before irreversible actions
- LLM agents that need human approval before executing tools
- Infrastructure automation with audit requirements
- Engineers integrating LLM or rule-based agents with production systems
- SRE/Platform teams adding guardrails to automation (CI/CD, infrastructure, IoT)
- Security and risk teams needing auditability and reversible changes
Requirements: Python 3.10+, no external packages.
git clone https://github.com/mfrethy-oneandall/stewardship-layer.git
cd stewardship-layer
python -m venv .venv
source .venv/bin/activate
python REFERENCE_IMPL/python/cli_demo.py
python -m unittest discover -s REFERENCE_IMPL/python/tests -v- STEWARD.md — Behavior contract and five-step loop
- SPEC.md — Full specification (actors, interfaces, state machine, threat model)
- PATTERNS.md — Design patterns for stewardship
- AGENT_SUMMARY.md — Quick reference for agent integration
- schemas/ — JSON Schema definitions for Proposal and Decision
See CONTRIBUTING.md for guidelines.
Maintained by Mike Frethy (@mfrethy-oneandall)