How do I verify a multi-agent workflow as one proof?
To verify a multi-agent workflow as one proof, you bind every hop of the task, the delegations between agents, and the final outcome into a single content-addressed reference, then recompute that reference offline. AlgoVoi Keystone Journey is the open, Apache-2.0 binding for this: it takes a whole multi-agent task and emits one journey_ref that verifies continuity, scope, and completeness from the bytes alone, with no AlgoVoi software in your trust base. It is published on PyPI as algovoi-journey-ref, byte-identical in Python and TypeScript.
Last updated 22 September 2026
When a crew, a graph, or an agent-to-agent task runs, the evidence is usually scattered across per-step logs that nobody can check as a whole. A journey reference collapses that into one proof: which agents acted, in what order, what each was authorized to do, and how the task ended. Anyone can recompute it and see that no hop was added, dropped, or altered.
What a journey reference binds
A journey_ref is computed as "sha256:" + SHA-256(JCS(fields)) over the ordered task, so it is tamper-evident and recomputable:
- Every hop, in order, as the task moved between agents.
- The delegations between agents, so scope narrowing is visible.
- The task's outcome, bound into the same reference.
Because the reference is content-addressed, omitting a hop or widening a delegation changes the bytes and fails recomputation. The proof does not depend on trusting the operator that produced it.
Where it fits with the rest of Keystone
Journey is one of the open cross-party bindings that compose agent hand-offs as offline-verifiable Keystone references. It sits alongside algovoi-revocation-ref, a chainable, offline-verified revocation envelope, and algovoi-delegation-ref, a tamper-evident cross-party delegation reference with depth and scope narrowing. Each is Apache-2.0 on PyPI and public GitHub. The aggregation proofs, the framework adapters, and Command Center ingestion are the commercial tier.
Verify offline, in either language
Python and TypeScript produce the same journey_ref byte for byte, so a proof built in one language verifies in the other. You recompute the reference from the recorded task and compare, no network call and no AlgoVoi service required. That is what makes a multi-agent run auditable by a counterparty, a customer, or a regulator who was not there when it ran.
Get started
Install the open binding from PyPI at https://pypi.org/project/algovoi-journey-ref/, and read how journeys, delegations, and revocations compose into a verifiable trust chain in the keystone guide at https://docs.algovoi.co.uk/keystone.