Skip to content

0001 — Open-Standard Agent Collaboration Platform

Status: Accepted

Context

kagent (agents on Kubernetes) and agentgateway (the AI data plane) are strong at creating and governing AI assets — LLMs, MCP tools, A2A agents — but neither ships a good human↔agent collaboration surface. The question this project answers: what is the right UI/fabric for humans and agents to co-inhabit, address each other by @mention, and delegate work?

The honest 80/20 framing (recorded so nobody has to re-litigate it):

  1. For a single solo developer who only wants "let me chat with my agents," standing up a full homeserver is over-engineering. kagent already ships Slack/Discord/Telegram→A2A bridge examples that deliver that UX with zero homeserver — that is the correct answer for that case, and Matrix is genuinely over-built for it.
  2. A proprietary SaaS (Slack/Discord/Teams) is the fastest path to a chat UI, but bakes vendor lock-in, foreign data residency, and per-seat cost into the critical path — disqualifying for a sovereignty showcase.

Matrix (and the wider open-standard stack) earns its keep only when three requirements are simultaneously real: (a) multiple humans, (b) self-hosted / sovereign operation, and (c) multi-network interoperability.

Decision

Build the platform entirely on open, mature, governed standards, and treat it as the reference build for exactly the (a)+(b)+(c) case above:

  1. Matrix as the collaboration fabric + UI (ADR 0002) — an IETF-track, foundation-governed protocol already run by NATO, the French and German governments, and Sweden's eSam.
  2. A2A (Linux Foundation) as the agent-to-agent delegation protocol (ADR 0004).
  3. MCP (Linux Foundation / Agentic AI Foundation) for agent tool access.
  4. Kubernetes + OIDC (via MAS, ADR 0003) as the runtime and identity substrate.

There is no proprietary SaaS in the loop and every layer is swappable. Unlike the sibling dev.fmind (a cost-capped free-tier mini-cloud), Fgentic deliberately does not optimise for billing or a single tiny node — it optimises for demonstrating the complete, enterprise-credible pattern, verified against upstream HEAD on 2026-07-08.

Consequences

  1. The platform is a genuine de-risking artifact for enterprises: every hop is currently-shipping OSS with no vendor to be captured by, and every fork chooses the enterprise-credible option.
  2. The cost is real operational weight (a homeserver, an auth service, a bridge) that a solo "just chat" use case would not justify — accepted, because that weight is the thing being demonstrated.
  3. Open standards buy interoperability for free later (Matrix bridges to WhatsApp/Slack/Signal/…), a payoff no proprietary UI can match.
  4. Provider independence is a hard constraint: workloads stay plain Kubernetes; only an optional Terraform GKE reference is cloud-specific.