Docs menu
Docs

Documentation

Concepts and short guides for the Agent Action Capsule — the open profile for verifiable records of what an AI agent did, and the transparency layer it builds on.

Concepts

You've reached a witness
Followed a link from a checkpoint or a receipt? Start here — what it shows, what it never does, and how to check it yourself.
What is an Agent Action Capsule?
A signed, tamper-evident record of an agent action — and the three properties that make it verifiable.
How it works
Both sides: the organisation seals each action and registers a checkpoint with a Transparency Service, which fixes the date; later, a party entitled to ask requests a slice and recomputes the answer offline.
Statement vs transparency layer
What happened vs where it is recorded — and why the statement is structure-independent.
What is a Transparency Service?
Register, receipt, log — and the boundary between a transparency service and a verifier.
Verifiable Data Structures
RFC9162_SHA256 (vds=1) vs CCF ccf.v1 (vds=2), and what stays constant across them.
How verification works
Three independent checks: the capsule_id, the signature, then the inclusion proof — all from the bytes, offline.
Trust map: what each check shows
Each check in plain words, for any exchange between agents: what it shows, what it does not show, and how it reads in two settings.
What's consequential
Seal what changes the world, plus reads of sensitive data; everything else is observability.
How it composes with your stack
Identity, authorization and transparency logs stay where they are; the capsule references their evidence by digest.
Witness anywhere
Only a digest and a timestamp leave: witness at the public log, run your own, or self-host where the hash never leaves.
Building on it? Implementation, tutorials, and adapter guides live in the canonical capsule-emit docs: capsule-emit/docs ↗. These pages cover the standard-level concepts; the repo docs cover hands-on usage.
Status. The Agent Action Capsule profile (draft-mih-scitt-agent-action-capsule, latest revision on the Datatracker) is an individual IETF Internet-Draft — submitted for discussion, not adopted by a working group, and not a standard. A companion draft, the Canonical Payload Binding (draft-mih-sokolov-scitt-payload-binding, latest revision on the Datatracker; site), defines the shared canonicalization and digest-binding layer. Both build on the IETF SCITT Architecture (RFC 9943) and the COSE Receipts specification (RFC 9942, now published).