Architecture Overview¶
HART OS (Hevolve Hive Agentic Runtime) is crowdsourced compute infrastructure for autonomous Hive AI Training -a distributed multi-agent platform where no single entity monopolizes AI. The diagram below shows the high-level component relationships.
System Diagram¶
┌─────────────────────────┐
│ External Clients │
│ (Discord, Telegram, │
│ Slack, Matrix, React) │
└───────────┬─────────────┘
│
▼
┌─────────────────────────┐
│ Flask Server (:6777) │
│ hart_intelligence_entry.py │
│ │
│ /chat /status /a2a │
│ /api/settings/compute │
│ /api/revenue/dashboard │
│ /social/* /coding/* │
└──┬───┬───┬───┬───┬──────┘
│ │ │ │ │
┌────────────────┘ │ │ │ └────────────────┐
▼ ▼ │ ▼ ▼
┌──────────────────┐ ┌────────┴───┴────────┐ ┌──────────────────┐
│ Recipe Pipeline │ │ Agent Engine │ │ Social Platform │
│ │ │ │ │ │
│ create_recipe │ │ goal_manager │ │ 82+ endpoints │
│ reuse_recipe │ │ dispatch │ │ communities │
│ helper │ │ speculative_disp │ │ posts/comments │
│ lifecycle_hooks │ │ revenue_aggregator │ │ karma/gamify │
└──────┬───────────┘ │ budget_gate │ │ encounters │
│ │ federated_agg │ └──────┬───────────┘
│ └──────────┬──────────┘ │
│ │ │
└────────────┬────────────┘ │
▼ │
┌──────────────────┐ │
│ SmartLedger │◄────────────────────────┘
│ agent_ledger │
│ │
│ JSON persistence│
│ Task tracking │
│ Cross-session │
└──────────────────┘
┌──────────────────┐ ┌──────────────────┐
│ Security Layer │ │ Service Tools │
│ │ │ │
│ master_key │ │ vram_manager │
│ hive_guardrails │ │ runtime_manager │
│ key_delegation │ │ model_lifecycle │
│ runtime_monitor │ │ whisper_tool │
│ node_watchdog │ └──────────────────┘
└──────────────────┘
┌──────────────────┐ ┌──────────────────┐
│ Peer Network │ │ Database │
│ │ │ │
│ gossip_protocol │ │ SQLite + WAL │
│ peer_discovery │ │ 60+ tables │
│ federation_sync │ │ hevolve_db.db │
└──────────────────┘ └──────────────────┘
Request Flow¶
- Client sends POST to
/chatwithuser_id,prompt_id,prompt - Flask server applies rate limiting, guardrail filtering, secret redaction
- Budget gate estimates LLM cost
- If
create_agent=trueor no existing recipe: CREATE mode - Decompose prompt into flows and actions
- Execute each action via LLM/tools
- Save recipe for future reuse
- If recipe exists: REUSE mode
- Load saved recipe
- Execute steps without LLM (90% faster)
- ActionState machine tracks each action's lifecycle
- SmartLedger persists state for cross-session recovery
- Response returned to client
Network Topology¶
Central (hevolve.ai)
│
├── Regional Host (us-east)
│ ├── Local Node (Nunba)
│ ├── Local Node (Nunba)
│ └── Local Node (Nunba)
│
├── Regional Host (eu-west)
│ ├── Local Node (Nunba)
│ └── Local Node (Nunba)
│
└── Flat Nodes (standalone)
Nodes discover each other via gossip protocol. Certificates are signed by parent tier. All messages are Ed25519-signed.
See Also¶
- task-delegation.md -- Task decomposition flow
- smart-ledger.md -- SmartLedger persistence
- federation-protocol.md -- Gossip protocol details