Context

Build the engine responsible for decomposing complex tasks into subtasks and routing them to the appropriate agents. This is a core orchestration capability that enables multi-agent collaboration on complex work.

§6.9 (Task Decomposability & Coordination Topology) adds task structure classification and per-task topology selection, which are key inputs to routing decisions. The auto topology selector uses task structure, tool count, and historical success rates to choose the optimal coordination approach.

Acceptance Criteria

Task Decomposition

• Parent task can be decomposed into multiple subtasks
• Subtask dependencies supported: sequential (A before B) and parallel (A and B simultaneously)
• LLM-based decomposition mode (agent decides how to break down task) → Deferred to Implement LLM-based decomposition strategy #168 (requires agent execution loop)
• Manual decomposition mode (via config or human input)
• Subtask progress tracking and rollup
• Parent task status derived from subtask statuses (e.g., all done → parent done, any failed → parent failed)

Task Structure Classification (§6.9 — new)

• task_structure field on task config: sequential, parallel, mixed
• Explicit classification — set by task creator or manager agent
• Inferred classification — derived from task properties (tool count, dependency graph, acceptance criteria structure)

Per-Task Coordination Topology (§6.9 — new)

• Per-task topology selection based on task structure and properties
• sequential → Single-agent (SAS) override
• parallel + structured → Centralized orchestrator
• parallel + exploratory → Decentralized peer debate
• mixed → Context-dependent (SAS backbone + delegation for parallel sub-tasks)

Auto Topology Selector (§6.9 — new)

• topology: "auto" config option triggers automatic selection
• Uses task_structure, tool count, and (when available) historical SAS success rate
• Configurable rules: sequential_override, parallel_default, mixed_default

Routing

• Auto-routing by agent skills and role matching
• Comprehensive unit tests (>80% coverage)

Dependencies

• Depends on hierarchical delegation (Implement hierarchical delegation (task flows down, results flow up) #12)
• Informed by §6.9 task decomposability research (Kim et al., 2025 — §16.3)

Design Spec Reference

• §6 — Task Workflow
• §6.9 — Task Decomposability & Coordination Topology (M4+)
• §16.3 — Agent Scaling Research