Key Takeaway
Require a short formal proof for every action an AI agent wants to take, and a small trusted checker enforces the contract so no protected action runs without that proof.
ON THIS PAGE
Key Findings
An enforcement layer called an Airlock accepts or rejects tool calls only after checking a machine-checkable proof that the call preserves the contract’s clauses. When contracts are written in a precise language and tool capabilities are modeled, the Airlock guarantees the contract’s clauses hold throughout execution for every agent and every sequence of requests under explicit assumptions. The approach is mechanized in Lean 4 and supports reusable capability modules (payments, document access, bookings) so common agreements can be composed and reused Capability Attestation Pattern.
Not sure where to start?Get personalized recommendations
Data Highlights
1100% of protected tool calls are blocked unless accompanied by a valid, machine-checked proof (no exceptions once deployed).
2Guarantees (commitment and completion) apply across 100% of possible executions for any agent, given the paper’s explicit assumptions about models and the environment.
31 formal mechanization: the design and main theorems are implemented and checked in Lean 4, with code and proofs published for audit and reuse.
What This Means
Engineers building agents that act on behalf of users or companies — especially for payments, document handling, or bookings — can use this to prevent unauthorized actions while still allowing autonomous operation. Technical leaders and platform operators can adopt an Airlock or capability modules to give clients verifiable assurances about what an agent may do. Capability Discovery Pattern
Key Figures

Fig 1: Figure 1: How PC2 works. Two parties sign a contract. The agent that does the work runs inside the Airlock and keeps its usual loop of reasoning and tool calls, but each tool call carries a proof that the contract holds throughout the operation. The Airlock checks the proof before the call runs. If the proof is correct, the call is performed and its result returned; if not, nothing runs, and the agent receives the reason, as an error message it can fix. The contract is settled when the agent proves its goal, and terminated otherwise, for instance at expiration, with a refund.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Writing precise, machine-checkable contracts and accurate models of tools is a nontrivial effort and may require templates or expert input. The guarantees hold under clear assumptions about how tools, hosts and external services behave; if those deployment assumptions fail, safety can break. Finding proofs for complex actions can be computationally expensive and may require automated theorem provers or specialized tooling, so runtime performance and proof search costs must be managed. This is related to broader design considerations captured by Evaluation-Driven Development (EDDOps).
Full Analysis
Agents must often act autonomously (send payments, release documents, book services), but free-form text promises are not enforceable. Enforce commitments by turning agreements into formal contracts and requiring a proof for each tool call. An Airlock checks each submitted proof before executing the call; if the proof type-checks against the contract, the call runs, otherwise it is blocked and the agent is informed. The obligation that a call must prove is derived automatically from the contract, published view, and tool models, allowing the agent to justify one action at a time rather than proving an entire plan in advance. Mutual Verification Pattern and Consensus-Based Decision Pattern support aligning multiple agents and stakeholders on executed actions. The construction includes capability modules (e.g., payment handling) so common services can ship with their models and guarantees. A mechanization in Lean 4 is provided so parties can audit the proofs and compose verified modules, enabling deployments where third-party Airlocks enforce contracts even when agents cannot be inspected directly.
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
One author (Alessandro Rudi) has a strong h‑index (~28) indicating an established researcher; although it’s an arXiv preprint, the author reputation supports high credibility.