06 — Subagents

What Was Researched

Patterns for spawning, managing, and coordinating child agents from a parent agent — enabling parallel workstreams, task delegation, and specialized sub-task execution.

Which Sources Were Used

Key Findings

Hermes Delegation System (Most Advanced)

Hermes has the most sophisticated subagent system studied:

• delegate_tool.py (140KB) — Full delegation system for spawning isolated subagents
• async_delegation.py (22KB) — Async subagent orchestration
• Parallel workstreams — Multiple subagents can work simultaneously on different tasks
• RPC-based tool calls — Python scripts call tools via RPC, collapsing multi-step pipelines into zero-context-cost turns
• Isolated contexts — Each subagent has its own conversation context (doesn't pollute parent's prompt cache)
• Iteration budget sharing — Parent shares its remaining budget with subagents
• Kanban plugin (plugins/kanban/) — Multi-agent board dispatcher + worker pattern for structured delegation

LangGraph Subgraphs

• Composable graphs — A graph can contain other graphs as nodes
• State isolation — Each subgraph maintains its own state
• Communication via state — Parent passes state to subgraph, receives modified state back
• Branching — Conditional routing to different subgraphs based on state

OpenClaw Multi-Agent Routing

• Agent routing — Route inbound channels/accounts to isolated agents
• Per-agent sessions — Each agent has its own workspace and session history
• Session tools — sessions_list, sessions_history, sessions_send for inter-agent communication
• Session spawning — sessions_spawn creates new agent sessions

Key Design Decisions

What Is Confirmed

1. Context isolation is critical — Subagent conversation context must not pollute parent's prompt cache
2. Budget sharing prevents runaway costs — Parent must limit subagent token consumption
3. RPC-based tool access is elegant — subagents call parent's tools without duplicating registrations
4. Kanban/task-board pattern (Hermes) enables structured multi-agent workflows
5. Multi-model deliberation achieves beyond-frontier performance — Budget fusion panels outperform individual frontier models on DRACO benchmarks CLAIM-157
6. Panel isolation prevents anchoring bias — Models must deliberate independently before cross-pollination CLAIM-145
7. Anonymity in peer review prevents lab-bias — Strip model identifiers to prevent self-preference CLAIM-150

What Is Uncertain

• Optimal number of concurrent subagents (cost vs. speed tradeoff)
• How to handle subagent failures gracefully (retry vs. escalate to parent)
• Whether subagents should share memory with parent
• Optimal panel size for fusion (diminishing returns beyond 3–5 panelists observed)
• Whether adaptive stopping thresholds generalize across domain-specific tasks

How This Applies to Building a Modern Model-Agnostic Agent Harness

1. Implement subagent spawning with context isolation — separate conversation per subagent
2. Share budget between parent and subagents — prevent cost explosions
3. Support parallel execution — multiple subagents working simultaneously
4. RPC-based tool sharing — subagents access parent's tools via RPC
5. Consider task-board orchestration (Hermes Kanban) for complex multi-agent workflows
6. Implement gateway-level fusion tool — Panel + Judge deliberation as a reusable tool CLAIM-145
7. Use structured JSON judge schemas — consensus, contradictions, blind spots, unique insights CLAIM-145
8. Support multiple deliberation patterns — Fusion for research, Council for debate, Swarm for task decomposition

Multi-Model Deliberation Patterns (Detailed Research)

For comprehensive research on multi-model orchestration patterns — including taxonomy, self-hosted implementation, anti-patterns, benchmarks, and decision matrices — see:

→ multi_model_deliberation_and_swarms.md

Patterns covered: | Pattern | Architecture | Hallmark | | :--- | :--- | :--- | | Panel + Judge (Fusion) | Parallel fan-out → structured judge → synthesis | OpenRouter Fusion | | Mixture-of-Agents (MoA) | Multi-layered iterative refinement | Wang et al. (ICLR 2025) | | Council / Debate | Multi-round peer review with chairman | Karpathy's llm-council | | Supervisor-Worker Swarm | Hierarchical delegation to specialists | Hermes Team*, CrewAI | | Graph Orchestration | Explicit state machine with conditional edges | LangGraph StateGraph |