Skip to content

Scenario 4 — Federation-aware agents

Purpose

Validate that agents running their own local ai-memory replica (backed by the ship-gate-validated 3-peer federation mesh) behave identically to agents hitting a single authoritative store. The A2A promise must hold whether operators deploy centrally or decentrally.

Mechanics

  1. Reconfigure: in addition to node-4's authoritative serve, stand up a mini 3-peer federation on nodes 1/2/3 (each node's ai-memory runs in --quorum-writes 2 --quorum-peers <other two> mode) pointed at each other. node-4 stays as the "central" store for comparison.
  2. Agent A on node-1 writes 50 memories via its LOCAL ai-memory replica under quorum.
  3. Wait the ship-gate Phase 2 settle bound (90 s).
  4. Agent B on node-2 recalls via its LOCAL ai-memory replica.
  5. Compare Agent B's recall count against Agent A's write count and enforce ≥ 95% convergence per the Phase 2 pass criterion.
  6. Run quorum probes analogous to Phase 2: kill node-3's serve, confirm Agent A's next write still returns 201. Kill node-2's serve too, confirm the next write returns 503.

Pass criterion

  • Convergence ≥ 95% of ok writes on both surviving peers after 90 s settle.
  • Quorum probes classify correctly (one-peer-down → 201, both- peers-down → 503).

Report shape

{
  "scenario": 4,
  "pass": true,
  "writes_ok": 50,
  "converged_node1": 50,
  "converged_node2": 50,
  "probe_one_peer_down": "201",
  "probe_both_peers_down": "503",
  "reasons": [""]
}

What a green result proves

  • Agents can run local ai-memory replicas with correct A2A semantics. No single-point-of-failure required.
  • The ship-gate Phase 2 guarantees extend cleanly to the A2A layer — nothing in the agent framework breaks quorum.

What a red result would mean

  • A regression between the ship-gate (which passes the same test) and the A2A harness's federation configuration. Either the harness misconfigured quorum, or a subtle interaction with the agent framework's MCP transport changed behaviour.

For three audiences

This is the test that proves you don't need a central server for your agents to talk through ai-memory. If you deploy three agents each with its own local copy, they'll stay in sync automatically.

Deployment flexibility is a customer-facing feature. This scenario proves the same correctness guarantees hold in the decentralized topology as in the centralized one. A red here means either topology breaks; a green means both work.

Re-runs ship-gate Phase 2's 200-write / 95%-convergence / quorum-probe contract, but through real MCP clients under agent workloads rather than synthetic curl bursts. Catches divergence between the HTTP-direct Phase 2 story and the MCP-client-initiated A2A story.