- Status: accepted for 2.0 (to be revisited by the community); part 2 implemented 2026-10-03
- Date: 2026-09-26, revised 2026-09-27 and 2026-10-03
- Deciders: the author; open for review with the rest of the 2.0 branch
Context
The question was whether Aegis should also govern the Agent2Agent protocol (A2A), the Linux
Foundation project for agents that discover and delegate work to other agents
(specification). Aegis 2.0 governs MCP: an agent
calling tools, reading resources and fetching prompts on systems such as Spark, Flink and Kafka.
What the A2A specification defines, in summary (paraphrased from the specification):
- an Agent Card that describes an agent’s identity, skills, endpoint and authentication;
- operations such as SendMessage, SendStreamingMessage, GetTask, ListTasks, CancelTask,
SubscribeToTask and push-notification configuration, over JSON-RPC, gRPC or HTTP+JSON bindings;
- stateful tasks with a lifecycle (working, input required, completed, failed, canceled, rejected);
- push notifications: the remote agent POSTs task updates to a webhook URL the client supplies.
Decision
Two parts, and the distinction matters more than the original version of this ADR admitted.
- Aegis 2.0 does not speak A2A on its front door. There is no A2A endpoint, no Agent Card
served, no task lifecycle, no push-notification webhook. The reasons are below.
- Calling another agent is a tool call, and that is already the governed shape. A remote A2A
agent can be exposed to the model as an ordinary MCP tool through a
ToolSource, in which case
every existing control applies to the delegation: classification, resource scoping, policy,
egress, rate limits, approval, audit, output redaction and screening. That is the accepted
direction for A2A support, and it needs no new engine.
The original ADR only stated part 1, which left the impression that A2A traffic simply cannot be
governed by Aegis. That was wrong in an important way: the interesting risk is not “Aegis does not
serve A2A”, it is “a model delegates work to another agent and nobody recorded or bounded it”. Part 2
is where that risk is actually addressed: delegated work becomes an ordinary governed tool action
with an explicit caller, target, bounded resource surface and audit record.
Original design sketch, superseded by the implementation below
The following records the decision as it stood before implementation. It is not a description of the
current source:
- one source,
type: a2a, per remote agent; the sketch proposed making its Agent Card skills tool
definitions and pinning them like other definitions;
- the sketch proposed
MUTATE as a per-skill default;
- the sketch proposed
SendMessage with polling to completion and no subscriptions or webhooks;
- the agent would see a gateway credential; the gateway audit record would retain the inbound caller
identity and any delegation chain that the own host or Knox had already established.
Update 2026-10-03: part 2 is implemented, with four changes from the sketch above
The a2a source now exists (aegis-source-a2a,
profile). Part 1 is unchanged: there is still no A2A
front door. Three points differ from the sketch, each for a reason found while building it:
- One tool per agent, not one per skill. An A2A 1.0 message has no field that selects a skill;
the remote agent decides from the content. A per-skill tool would carry a classification the
protocol cannot enforce, so the agent endpoint is the governed unit and the skills are listed in its
description.
- Unclassified means
DESTRUCTIVE, like every other source, rather than a special MUTATE
default. The reference profile classifies its agent MUTATE explicitly.
- Protocol scope: discovery accepts advertised JSON-RPC Agent Card interfaces whose protocol version
begins with
1. and emits requests with A2A-Version: 1.0. The shipped evidence validates an A2A
1.0 agent only; accepting a later 1.x card is not an interoperability claim until it has its own
reviewed evidence.
- No polling. A call sends one
SendMessage and waits for the reply within the timeout. A task
that is not finished, or that asks for more input, is returned as a tool error; tasks are not
continued across calls.
The credential sent to the agent is the source’s own, or a per-caller token when the source uses
credentials.type: token-exchange. Evidence: A2aToolSourceTest, and A2aAgentFrontDoorsTest
against an agent built on the official a2a-sdk 1.2.1, through both front doors.
Reasons for keeping the front door MCP-only
- A different unit of governance. Aegis decides on a tool call: a named operation, typed
arguments, typed resources, a READ / MUTATE / DESTRUCTIVE class, an exact argument digest that an
approval is bound to. An A2A message is natural-language or structured content addressed to
another agent, whose effect on real systems is decided by that agent, not visible to the
gateway. Several controls that make Aegis useful (resource extraction, dry run, a receipt for the
exact action) have little to bind to. Exposing the agent as a tool does not remove that limit; it
makes it explicit, which is why an operator must classify the agent by its strongest possible effect.
An unclassified A2A tool is
DESTRUCTIVE; the reference profile explicitly classifies its inventory
agent MUTATE, and both classes require approval.
- A different state model. A2A tasks are long-lived, stateful objects with their own
lifecycle. Aegis 2.0 is designed so that no governance state depends on protocol state (that is
why it works for both the stateful 2025 MCP and the stateless 2026-07-28 MCP). Governing A2A
properly means tracking task ownership and every status and artifact update, which is a
different engine.
- A different direction of traffic. Push notifications are server-initiated HTTP calls to a
client-chosen URL. Aegis deliberately refuses server-to-client requests on MCP streams; accepting
webhooks would reverse a core safety property and needs its own egress and SSRF design.
- Scope creep would weaken what exists. A half-built A2A proxy would be a second, less tested
policy surface next to a well-tested one. The review of 2.0 is easier, and the result safer, if
the protocol surface stays exactly MCP.
What still works today
- An agent that is itself an A2A server can call its tools through Aegis over MCP. The actions
that touch real systems are then governed, whichever protocol brought the request to that agent.
- The inbound delegation chain is retained in Aegis’s caller identity and audit record: RFC 8693
act
claims at the own host and ActorChainPrincipal in Knox become CallerIdentity.actorChain, and
delegated identities cannot approve. The A2A source does not add an A2A actor to that chain.
- W3C trace context joins an A2A source call to the Aegis audit record when the upstream client forwards
traceparent.
Future work on the A2A decision
- A2A tool source (part 2): implemented for one A2A 1.x JSON-RPC Agent Card interface using
SendMessage, with evidence against an A2A 1.0 agent. Future work is an explicit interoperability
policy and evidence for later 1.x versions, card-signature verification, other bindings, streaming,
and task continuation. The implementation is documented in
adopters/configurations/a2a.
- A2A front door (part 1): requires A2A to define typed, side-effect-declaring skills so a message
maps to a classifiable action, and a community decision to add a second protocol engine with its own
tests rather than stretching this one.
Consequences
- 2.0 has an A2A client (the
a2a source, part 2) and no A2A server. This ADR should not be read as
claiming that Aegis governs A2A traffic that does not pass through that source.
- The engine SPI (
PolicyDecisionPoint, AuditSink, StateStore, DecisionListener) is
protocol-neutral, so a future A2A front door could reuse policy, audit and state without
changing the MCP path.