Anemoi A2A Research Synthesis

Date: 2025-12-30 Source: Perplexity Research (CONTENT.md)

Key Findings

1. Anemoi Architecture

Thread-based communication via MCP server with 5 primitives: ``` list_agents() → Discover available agents create_thread(P_0) → Initialize conversation context send_message(τ, msg) → Broadcast/directed message wait_for_mentions(τ) → Block until mentioned (real-time awareness) close_thread(τ) → Terminate with outcome logging ```

Performance: +9.09% on GAIA benchmark vs centralized approaches (OWL, CAMEL, AutoGen)

2. Scenario-Pattern Mapping

Scenario	Best Pattern	Effort	Why
A: Fork Context	Agent Continuations	1-2 wks	JSON state blob transfers full context
B: Multi-Session	Anemoi A2A	3-4 wks	Real-time `wait_for_mentions` beats Redis polling
C: Pre-Launch	Hybrid	2-3 wks	Continuations + A2A negotiation
D: Hook Chain	Anemoi (optional)	1 wk	Message instrumentation on existing hooks

3. Trade-Off Matrix

Dimension	File-Based (Now)	Continuations	Anemoi A2A
Latency	50-200ms	0ms (memory)	5-20ms
Multi-Session	Eventual	Manual poll	Real-time
Token Overhead	High	Very Low	Very Low
CLI Support	Native	Native	Needs wrapper
Effort	0	1-2 weeks	3-4 weeks

Recommendation from Perplexity

Phase 1: Agent Continuations (Immediate)

Replace CONTEXT.md with continuation_state.json
Fork operations pass state blob, not empty context
Solves: Context loss on fork, pre-launch isolation

Phase 2: Anemoi A2A (Parallel if capacity)

Deploy Anemoi MCP server
Replace Redis pub/sub with thread messaging
Solves: Multi-session coordination delay

Phase 3: Semi-Centralized (Defer)

Full architecture redesign
Wait until Phases 1-2 validated

Decision Required

Question: Which pain point is most urgent?

If...	Then...
Fork context loss is blocking work	Start Phase 1 immediately
Running 3+ parallel sessions often	Prioritize Phase 2
Current system "good enough"	Document patterns, defer implementation

Self-Inquiry Checklist (From Research)

Before implementing, answer:

Context loss on fork: What % of parent state is needed by child?
- <20% → File-based sufficient
- 50%+ → Continuations essential
Token waste: What % goes to redundant context injection?
- <5% → Redis webhooks acceptable
- >15% → Anemoi threading worthwhile
Pre-launch clarification: How often does main session need more from pre-launch?
- Rarely → One-shot acceptable
- >30% → Bidirectional negotiation valuable
MCP server maturity: Already running iaip/coaia MCPs?
- Yes → A2A server is low-friction
- No → Requires SDK setup (2-3 days)
Shell script integration: Must LAUNCH__*.sh stay in loop?
- Yes → Anemoi needs API wrapper
- No → Native support available

Implementation Artifacts Needed

For Phase 1 (Continuations)

Modify newsessionuuid to generate continuation JSON
Update miette_newsession_context.sh to write continuation (not CONTEXT.md)
Add --context flag handling to launcher scripts
Test fork/resume with state preservation

For Phase 2 (Anemoi A2A)

Deploy Anemoi MCP server (or reference implementation)
Add to .mcp-config as anemoi-a2a
Replace Redis webhook with send_message(τ, ...)
Add thread instrumentation to v2_claude_hooks
Test multi-session scenarios

References

Anemoi: https://github.com/Coral-Protocol/Anemoi (arXiv:2508.17068)
Agent Continuations: https://github.com/SnapLogic/agent-continuations
SagaLLM: arXiv:2503.11951 / VLDB 2025
MCP-Zero: arXiv:2506.01056