Concord
Contracts for the actions AI agents take.
Durable execution tells you what ran. Concord records what the action meant, who authorized it, which systems it touched, and how to undo it when something goes sideways.
Oversight of agent actions is an openly unsolved problem, by the published account of research groups inside the companies shipping agents. Concord is a proposal for the missing layer.
Concord is the action layer of a larger governance stack. How it fits with Lattice: The worker and the action.
One command shape across agent loops, deterministic jobs, connector calls, approvals, artifacts, memory, and compensation.
What is an agent contract?
An agent contract is a record attached to each action an AI agent takes: what the action meant, who authorized it, which systems it touched, and how to undo it. Durable execution preserves the running of work; an agent contract preserves its meaning.
The term comes from Avianna’s Concord project. Where an employment contract governs a worker, an agent contract governs a single action — proposed, authorized, executed, recorded.
Read this in one sitting
Why agents need a contract
Durability preserves execution. It does not preserve intent. This piece frames the gap Concord closes.
Read part 1 02Contract, not runtime
Draws the boundary between Concord, DBOS, LangGraph, OPA, OTel, and Postgres with interactive demos.
Read part 2 03A day in the life
Traces a hotel booking end to end, then toggles between full Concord, minimal Concord, and no contract.
Read part 3Architecture and worked examples
See the primitives under pressure
Agent contracts, briefly
Is Concord a runtime?
No. Concord is a contract layer that sits above durable runtimes like DBOS or Temporal. The runtime guarantees the work runs to completion; the contract records what the work meant, who allowed it, and how to reverse it.
How is an agent contract different from durable execution?
Durable execution preserves state and retries so a workflow survives crashes — it tells you what ran. An agent contract adds the layer audits actually need: intent, authority, blast radius, and a compensation path.
How is Concord different from a policy engine like OPA?
A policy engine answers “is this allowed?” at decision time and forgets. Concord enforces that check on every action and keeps the ruling — meaning, authority, undo — in a ledger that later trust decisions and audits can read.
Who needs agent contracts?
Any team letting agents act on real systems: booking, refunding, updating records. The moment an agent proposes work instead of a human, execution history stops being an explanation — that is the gap the contract closes.