**To:** oauth@ietf.org
**From:** Christian Nennemann <ietf@nennemann.de>
**Subject:** draft-oauth-transaction-tokens-for-agents-06: complementary work on DAG-based delegation (draft-nennemann-act)

Hi all,

I noticed the publication of draft-oauth-transaction-tokens-for-agents-06
(Raut et al., 2026-04-11) and wanted to share some complementary work that
addresses an adjacent slice of the agent-delegation problem space. The
Amazon draft fills a real gap at the OAuth authorization-server layer, and
I think there is useful coordination potential rather than overlap.

# Technical difference in one paragraph

draft-oauth-transaction-tokens-for-agents introduces `actchain` as an
ordered array documenting delegation history, plus `agentic_ctx` carrying
type/version/intent/operational constraints, with a split between
principal-initiated and autonomous flow types. Our work
(draft-nennemann-act-01) models delegation history as a DAG through a
`pred` (predecessor) claim that is itself an array of parent token
references. A linear `actchain` is a special case of the DAG form where
every node has exactly one predecessor.

# Why a DAG, concretely

Consider an agent that fans out to N parallel sub-agents (e.g. one per
data source) and then synthesizes a single response from their results.
The synthesis step has N predecessors, not one. A linear `actchain`
cannot express this fan-in; you would have to either linearize artificially
(losing causality) or emit N parallel chains (losing the join). With a
DAG-valued `pred`, the synthesis token references all N predecessor tokens
directly, and a verifier can walk the graph to check that each parallel
branch was authorized and unexpired. Fork, join, and diamond topologies
fall out of the same structure.

# Layering, not competition

These two drafts sit at different layers:

- Txn-Tokens-for-Agents is anchored at an OAuth authorization server:
  the AS mints and validates tokens, and `actchain` is read in the
  context of an AS-issued transaction token.
- ACT is designed for peer-to-peer agent orchestration without
  requiring an AS in the hot path — useful for multi-vendor agent
  meshes where no single AS is authoritative. It is transport-agnostic
  and leans on JWS for provenance.

An AS-issued txn-token could carry an ACT-shaped `pred` graph
internally, or an ACT chain could terminate at an AS that upgrades it
into a txn-token for a specific resource. The two seem stackable.

# Offer

Happy to compare test vectors, especially around:

- claim naming: `agentic_ctx` (Raut) vs ACT's `task` claim — is there
  an opportunity to align on a shared intent/constraint shape so
  downstream verifiers don't have to parse both?
- linear-subset interop: confirming that a degenerate DAG (each node
  one parent) round-trips cleanly to/from `actchain`.
- autonomous-flow semantics: how ACT's unattended-delegation marker
  maps onto Raut's autonomous flow type.

ACT draft: https://datatracker.ietf.org/doc/draft-nennemann-act/

Feedback welcome, on- or off-list.

Best,
Christian Nennemann
Independent Researcher
ietf@nennemann.de