Research program
Proof Carrying Contracts (PC2)
An open protocol for enforceable agreements between AI agents acting for different people and organisations. Before an action covered by the agreement runs, the agent must supply a machine-checkable proof that it preserves the agreed constraints.
The problem
AI agents increasingly act for people and organisations: they buy services, exchange information and delegate work to other agents. What an agent must respect is written in instructions, policies or agreements, but text does not make it comply. It may make mistakes, be redirected by what it reads, or pursue interests different from those of the party it deals with.
An economy of such agents needs agreements that each party can rely on without inspecting or trusting the other party's agent.
The approach
- The parties agree on a contract: the task, the resources, the constraints that must hold and what counts as completion.
- Before an action covered by the contract runs, the agent must supply a machine-checkable proof that the action preserves the agreed constraints in every case allowed by its model. An execution environment, the Airlock, checks the proof and controls access to the action.
- The parties specify what must be preserved; the agent supplies the reasoning that justifies its decisions and remains free to adapt its strategy. No separate permission policy has to be designed and proven safe.
- Delegation is checked in the same way: a subcontract is admitted only if it fits within the parent contract, so commitments are preserved when work passes to other agents.
The paper proves that the agreed constraints hold at every instant of execution, even for an adversarial agent, under stated assumptions, and identifies when this implies safety in the world. The research programme now studies whether useful requirements can be formalized faithfully, whether agents can act under them at practical cost, and what protection the resulting system provides.
What is available
- Paper
- Proof-Carrying Contracts: Commitments AI Agents Cannot Break. C. Ciliberto, A. Rudi, E. Tassi. arXiv:2610.28484, October 2026, 70 pages.
- Website
- The PC2 website: an overview of the protocol, a narrated walkthrough of a PC2 interaction and the research programme.
- Code
- github.com/consequential-ai/proof-carrying-contracts. The mechanization and code accompanying the paper are coming soon; a demonstration will follow.
People
- Carlo Ciliberto
- UCL; AI4I, Turin (PRAXIS)
- Alessandro Rudi
- SDA Bocconi, Milan (REAL)
- Enrico Tassi
- Inria (STAMP)
Take part
Seven research and engineering positions are being prepared, based in Milan (Bocconi, REAL) and Turin (AI4I, PRAXIS): see the positions. Discussion of the paper and its assumptions is welcome now; contribution instructions will come with the first implementation release.
Contact: alessandro.rudi@sdabocconi.it
Cite
@misc{ciliberto2026proofcarryingcontracts,
title = {Proof-Carrying Contracts: Commitments {AI} Agents Cannot Break},
author = {Carlo Ciliberto and Alessandro Rudi and Enrico Tassi},
year = {2026},
eprint = {2610.28484},
archivePrefix = {arXiv},
url = {https://arxiv.org/pdf/2610.28484}
}