The core idea: agents shouldn't operate software step by step. They should publish objectives, compare competing provider offers, accept the best one under policy, and let providers fulfill outcomes with evidence and settlement.
Why this matters: we're still building agents as "software operators" — better interns that click through UIs and call APIs. That forces probabilistic models to produce deterministic side effects. The failure modes (wrong clicks, prompt injection, "probably right" becoming "committed") keep piling up.
https://github.com/csehammad/covenant-layer
Covenant Layer moves the edge from procedures to commitments. The agent handles intent, comparison, approval. The provider handles execution, evidence, accountability.
What's in the repo: - Protocol spec - Rust + Solidity implementation (onboarding, conformance, registry) - E2E demo: JFK→CDG flight booking with 3 competing providers on Base Sepolia - Architecture diagrams and security rationale
Early draft. Would love feedback on the model, weak assumptions, or missing pieces.