ADR-0011.

Date: 2026-07-09 Status: accepted Deciders: Thibault Souris (maintainer) Method: decision-loop — ecosystem veille (2026), sourced state of the art, challenge, durable position

Context

The veille showed the neutral substrate of the agent ecosystem is now institutionalised under the Linux Foundation / Agentic AI Foundation and adjacent bodies: MCP (tools/context), A2A (agent-to-agent), OpenTelemetry GenAI semantic conventions (observability), and a converging agent-identity stack (OAuth 2.1/PKCE, SPIFFE/WIMSE, Entra Agent ID, IETF agent_assertion). For runward to stay neutral and current, the reference stack must name these standards as the default contracts of specific ports, resolved by an adapter decision — without runward ever exporting telemetry, providing a runtime, or bundling an SDK. These specs also churn (MCP revises roughly every eight months), so each must be isolated behind a versioned contract and re-tested on revision.

Decision

The reference stack names, per port, the neutral standard as the default contract, each resolved by a versioned adapter decision and pinned in an ADR:

• Tools / context port → MCP (pinned spec version; per ADR-0009, a consumed server is pinned by version and hash).
• Observability port → OpenTelemetry GenAI semantic conventions (spans invoke_agent, execute_tool; pinned — the spec is still experimental). runward prescribes the port and the span contract; the operator plugs the backend (CloudWatch GenAI, Langfuse, Phoenix, Braintrust…) as an adapter. runward exports nothing itself.
• Identity & delegation port → OAuth 2.1/PKCE + SPIFFE / Entra Agent ID / IETF agent_assertion, sober default: one identity per agent, short scoped credentials, explicit ownership and lifecycle; the decision matrix documents when to harden, on an objective trigger, in an ADR.
• Distributed-topology toggle → A2A, deferred default: no A2A until a single agent no longer suffices (a decision-matrix trigger, not a starting default).

runward references and pins; it never implements the standard, exports the telemetry, provides the runtime, or ships the sandbox/memory. Each standard's version is pinned in an ADR and re-tested on revision.

Alternatives discarded

• Adopt a vendor's proprietary equivalent (a cloud's native gateway, identity, or eval SDK). Breaks the neutrality that is the moat.
• Ignore the standards. runward drifts into irrelevance as the ecosystem converges around them.
• Implement a standard (ship an MCP server, an OTel exporter, an identity broker). That crosses the line into being a runtime — the one thing runward must never become.

Consequences

• Positive. runward stays neutral and current: the model, the runtime, the telemetry backend, the identity broker are all adapter decisions behind a port with a named, versioned contract. Isolation behind a versioned contract absorbs spec churn.
• Negative, accepted. A pinning-and-re-test discipline: each spec revision (notably MCP) forces a re-test and a new pinned version in an ADR.
• On other boundaries. reference-stack.md and decision-matrix.md name the standards as ports and add the identity-hardening and A2A/distributed triggers; the doctrine's topology and integration sections cite them. The zero-LLM gate is untouched.

Reevaluation trigger (mandatory, dated)

Reopen on each material spec revision — the finalized MCP 2026-07-28 first — to re-test the adapter and pin the new version; or when a listed standard is superseded by a more widely-adopted neutral one.

Trigger set on: 2026-07-09 · Watched via: MCP / A2A / OTel GenAI / agent-identity release feeds (the veille watch-list).

References

• MCP · A2A · AGENTS.md (Linux Foundation / Agentic AI Foundation); OpenTelemetry GenAI semantic conventions; OAuth 2.1 · SPIFFE · Entra Agent ID · IETF agent_assertion.
• ADR-0009 — the MCP version+hash pinning rule.
• templates/mission/reference-stack.md, templates/mission/decision-matrix.md — the surfaces this ADR touches.