Pact: A Choreographic Language for Agentic Ecosystems
New choreographic language embeds game theory into protocol design for trustless agent cooperation.
Recent advances in LLMs have spawned autonomous software agents that interact with untrusted counterparts and manage private data in open, multi-party settings. While choreographic programming offers correct-by-construction protocol design, it assumes cooperative participants—ignoring agent self-interest. To bridge this gap, researchers introduce Pact, a choreographic language that draws from game theory to describe agent choices and preferences. Every Pact protocol maps to a formal game, enabling protocol designers to reason about game-theoretic properties and solve for optimal decision policies.
Pact includes a preliminary implementation featuring a bounded-rational solver that computes decision policies over Pact protocols. The language was applied to multi-party coordination scenarios with self-interested agentic participants, demonstrating its ability to enforce cooperation even when agents have conflicting incentives. This work, to be presented at the 2nd International Workshop on Choreographic Programming (CP 2026), addresses a critical gap in building trustworthy agent ecosystems by ensuring protocols are not only correct but also incentive-compatible.
- Pact combines choreographic programming with game theory to model agent self-interest in multi-party protocols.
- Every Pact protocol maps to a formal game, allowing designers to solve for decision policies and reason about incentives.
- A bounded-rational solver computes optimal policies, tested on real multi-party coordination among self-interested AI agents.
Why It Matters
Enables trustless, reliable multi-agent cooperation in open ecosystems where agents have competing interests.