Hat Orchestrator¶

HAT Orchestrator — multi-party human-AI teaming component.

Cluster: Goal & Planning | Type: component | MCP Tools: None

Overview¶

Human-AI Teaming (HAT) orchestrator that manages a registry of heterogeneous agents (human, AI, robot) and coordinates them through task allocation, synchronisation points, authority resolution, and a Common Operational Picture (COP). Agents are registered with capability profiles and Sheridan authority levels; tasks are matched to the best-available agent; alerts can be broadcast to the whole team.

When to use:

Coordinating mixed human-AI-robot teams where authority and capability must be tracked per agent
Implementing synchronisation barriers that wait for all required team members before proceeding
Resolving authority conflicts between overlapping agents using Sheridan-level priority rules

Example:

from mvp.hat_orchestrator import HATOrchestratorBlock, HATOrchestratorInput

block = HATOrchestratorBlock(name="hat")
block.infer(HATOrchestratorInput(
    op="register_agent", agent_id="human-1", agent_type="human",
    capabilities={"review": True}, authority_level=8,
))
result = block.infer(HATOrchestratorInput(
    op="allocate_task", task_type="review", required_capabilities={"review": True},
))
# result.ok → True; result.value.allocation contains assigned agent

Works well with: autonomy_governor, autonomous_orchestrator, human_development

Operational Caveats¶

Pilot-ready orchestration, not a full task system

hat_orchestrator is suitable for local pilots and launch-plan workflows where a small number of agents need durable registration, task assignment, synchronisation barriers, authority resolution, and a shared Common Operational Picture. It persists state to SQLite and reserves automatic task IDs atomically, so stale block instances should not generate duplicate task_# IDs.

It now exposes complete_task and fail_task lifecycle operations, surfaces degraded allocation and persistence envelopes, and wires GuardrailConfig into registration and allocation chokepoints. Alerts remain caller-managed, but clear_alerts lets operators explicitly clear accumulated alert state. Agent availability is still manually supplied rather than derived from active workload, and authority/capability claims remain caller-supplied rather than externally identity-verified.

SQLite is appropriate for single-machine MCP/REST pilots. For high-concurrency, multi-host, or customer-critical deployments, use this component behind a single writer process or migrate the store to a server database with transactional task lifecycle operations.

Public API¶