Entity
Concordat (Typed Sessions for LLM Agents)
Concordat is a Python library that enforces agent actions through a finite-state protocol (drawing on typestate/session-type concepts) rather than relying solely on prompt instructions, aiming to guarantee procedural safeguards like mandatory customer confirmation before irreversible actions.
This page was written and is updated by the paper from the source articles listed below. Importance is how central the topic is to current coverage; confidence is how well-sourced the page is; mentions counts the articles that fed it. How the wiki works →
Concordat is a Python library designed to enforce procedural rules on consequential actions taken by LLM-based agents, such as cancelling bookings, issuing refunds, or changing accounts (arXiv, October 28, 2026). Rather than relying on prompt-based instructions that an agent only follows "as reliably as the model chooses to," Concordat treats an agent as an object whose interface is a finite-state protocol, drawing on concepts from typestate and session types in programming language theory, to enforce procedural constraints such as requiring explicit customer confirmation before an irreversible action (arXiv, October 28, 2026).
## Strategic Relevance
As enterprises deploy LLM agents with authority to take real-world consequential actions, mechanisms that move beyond prompt-level guardrails to structurally enforced protocols are of direct interest to legal and compliance teams assessing agentic AI risk. Typed-session or state-machine approaches to agent action governance represent a technical complement to policy- and identity-based AI agent governance frameworks, and may become a reference point in disputes over whether an agent's actions were properly authorized or auditable.