Skip to content

0022 — Evaluate AGNTCY (Identity, OASF, Directory) for Cross-Org Agent Identity and Discovery

Status: Proposed. Related: ADR 0021, ADR 0013, ADR 0015, ADR 0002, ADR 0010, issue #1431.

Context

Fgentic identifies and discovers remote agents across organizational boundaries with mechanisms it maintains itself:

  1. Remote A2A routes require a currently verified, operator-pinned ES256 Signed AgentCard, resolved out of band without a network fetch (ADR 0021). This is a deliberate, SSRF-safe trust anchor, but it is unique to Fgentic and carries no cross-vendor interoperability.
  2. Cross-protocol discovery reveals a remote org's Matrix and ActivityPub transports through one WebFinger lookup (fediverse.md). It answers "where and over what," not "what capabilities, dependencies, and status does that agent advertise in a shared schema."

The Linux Foundation's AGNTCY project (donated by Cisco, July 2025; "Internet of Agents") now ships vendor-neutral primitives aimed squarely at this problem space:

  1. AGNTCY Identity — onboard, create, and verify cryptographically verifiable identities for Agents, MCP servers, and multi-agent systems, intended for secure action across organizational boundaries.
  2. OASF (Open Agentic Schema Framework) — a standardized capability/dependency/status schema so any agent can describe and discover others.
  3. DIR (Distributed Announce and Discovery) — a decentralized Agent Directory built on OASF.
  4. SLIM — a low-latency agent messaging transport.

AGNTCY is a distinct Linux Foundation project from both the A2A project and the Agentic AI Foundation (AAIF, which hosts MCP, agentgateway, Goose, and AGENTS.md). open-stack.md currently maps AAIF, the A2A project, CNCF, and the Matrix.org Foundation, but not AGNTCY. Verified against primary sources on 2026-07-26: the Linux Foundation AGNTCY announcement and the AGNTCY GitHub organization (Identity, DIR, OASF, SLIM repositories).

Shared foundation membership is not a runtime boundary — the same principle applied to Goose in open-stack.md and to SPIFFE in ADR 0010. A capability schema or identity spec becomes a Fgentic control only when the deployed configuration can actually verify and enforce it end to end.

Decision

  1. Track, do not adopt. Record AGNTCY as a candidate for the federation identity (ADR 0021) and discovery (fediverse.md) seams, and evaluate it under the same evidence discipline as ADR 0010: current component versions and maturity reviewed against primary sources, mapped to concrete seams, with adoption gated on runtime evidence rather than membership or promise.
  2. Reject SLIM as a transport. Matrix is the human↔agent collaboration fabric (ADR 0002) and A2A is the delegation contract (ADR 0004); a competing agent transport is out of scope. Only AGNTCY's identity and discovery layers are candidates, and only as complements to — not replacements of — the Matrix/A2A critical path.
  3. Add AGNTCY to the governance map. When this ADR moves beyond Proposed, extend open-stack.md so the map distinguishes AGNTCY (LF) from the A2A project (LF) and AAIF (LF), with the same per-boundary claim limits.
  4. Do not weaken existing anchors while evaluating. The pinned-AgentCard trust anchor (ADR 0021), the unpinned-actor SSRF guard, and the federation lab acceptance rig (ADR 0013) remain authoritative until an AGNTCY-based replacement passes equivalent positive and negative acceptance.

Adoption triggers

Reopen this ADR toward Accepted only when all mandatory triggers hold:

  1. Stable surface: at least one AGNTCY identity/discovery component reaches a non-prerelease, versioned release whose spec is implementable without tracking a moving main.
  2. Seam fit: AGNTCY Identity can express and verify the exact remote-agent trust decision Fgentic makes today (a specific signer for a specific ghost/target), without a network fetch that reintroduces the SSRF surface ADR 0021 closes, or with a fetch whose risk is separately mitigated and tested.
  3. Two-sided value: a second, independently operated org (or the federation lab standing in for one) can publish and consume an OASF/DIR capability record such that discovery produces a decision Fgentic could not make from WebFinger + AgentCard alone.
  4. No sovereignty regression: discovery and identity resolution require no proprietary SaaS, no hosted directory Fgentic cannot self-run, and no dependency that breaks the closed-federation allowlist model.

Strengthening but not sufficient alone: more supporting vendors, ecosystem momentum, or a desire for a "standard" identity format.

Required adoption prototype

When triggers are met, acceptance requires a disposable federation-lab prototype (ADR 0013) that:

  1. Publishes an OASF/DIR record for a lab agent and resolves it from the peer org's side.
  2. Verifies an AGNTCY identity for a remote ghost and authorizes an A2A delegation on it; a wrong/absent/untrusted identity must fail closed, matching the current pinned-AgentCard negative tests.
  3. Proves the Matrix and A2A critical path is unchanged — AGNTCY sits alongside as identity/discovery, not as transport.
  4. Records what it replaces (pinning, WebFinger, or neither) and leaves the SSRF guard intact for unpinned/unknown actors.

Consequences

  1. The platform gains a tracked, sourced position on the leading Linux Foundation multi-agent identity/discovery effort without importing an unproven dependency or a competing transport.
  2. The federation identity story stays on the maintained pinned-AgentCard anchor until evidence justifies a change; no operational surface is added now.
  3. If AGNTCY stalls or fragments, this ADR is closed with the checked versions recorded — the same low-cost exit as ADR 0010. The decision is intentionally reversible in both directions.