Subagent Workflow

Definition

Subagent workflow is an agentic pattern where a foreground assistant delegates complex, long-running, specialized, or adversarial work to other agents and then integrates, reviews, or verifies their outputs.

Current Synthesis

Subagents remain a harness-level technique for preserving main-context clarity, expanding search, dividing roles, and cross-checking work. The pattern appears in coding, research, theorem proving, and productized multi-agent systems: one agent may explore code, another may test, another may argue against a plan, and a lead agent may synthesize the result.

The 2026 coding-agent discussion adds a model/harness co-evolution angle. Some workflows now need explicit leader, worker, and verifier roles created dynamically by the harness, but future models may internalize more of that decomposition. The practical conclusion remains unchanged: role boundaries, permissions, handoff documents, and verification make subagents useful rather than just parallel noise.

Key Claims

• Subagents preserve the foreground context when a task is too large, tool-heavy, or disruptive for the main conversation.
• Role-specific agents are useful when tasks need different permissions, viewpoints, or standards, such as code exploration, testing, adversarial review, or proof repair.
• Multi-agent cross-checking can correct drift and hallucination, but it consumes more tokens and still needs human or verifier acceptance.
• Handoff artifacts are necessary because a subagent’s output must be compact enough for another agent or human to reuse.
• Dynamic leader/worker/verifier teams are an emerging harness pattern for agent-native coding work.
• The durability of subagent orchestration depends on model/harness co-evolution: some decomposition may move into models, while permissions and verification remain external.

Evidence

• Background subagents and adversarial pro/con roles are described as reusable skill patterns for tool-heavy or high-token work: 阿里千问离职余震，在几万人的铁球里如何体面生存.
• Governance sources emphasize role-specific permissions, information boundaries, and handoff documents so agents do not overstep or repair tests dishonestly: 探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐.
• Multi-agent systems can exchange larger context than human feedback normally provides and can cross-check long-context drift: 当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent.
• Practical coding and theorem-proving sources show subagents used for planning, implementation, review, Lean proof attempts, and verifier-driven repair: Vol. 166 闲聊: 从 Gemini 到 AI 的加速与混沌, 137. 对洪乐潼的4小时访谈：AI for Math、把数学变成Lean、数学天书中的证明、直觉、被创造与被发现的.
• Agent-native coding tools are described as moving toward dynamically generated leader, worker, and verifier teams inside the harness: Ep 59. 2026 Agent 编程新趋势.

Counterevidence & Qualifications

Parallel agents do not automatically improve quality. They can multiply wrong assumptions, create integration work, burn tokens, and hide responsibility. Subagent workflows need explicit task boundaries, permission scoping, output contracts, and verification; otherwise the user receives more fluent uncertainty instead of a stronger result.

What Changed

• Dynamic multi-agent team formation is now included as a 2026 coding-harness pattern.
• The synthesis now ties subagent workflows to model/harness co-evolution, not only current Claude Code or MiniMax-style orchestration.
• Verification roles are now explicit alongside explorer, worker, critic, and synthesizer roles.

Related Concepts

• Agent Harness - orchestration layer that creates and constrains subagents.
• AI Skills - packaging mechanism for reusable subagent patterns.
• Multi-Agent Collaboration - broader frame for agents exchanging context and critique.
• Context Engineering - task-context design needed for handoff and integration.
• AI Coding Verification - acceptance layer that makes verifier agents meaningful.
• Model Harness Co-Evolution - question of which orchestration logic remains outside models.
• Agent Command Center - interface pattern where humans can supervise multiple agent sessions.
• AI For Math - theorem-proving domain where subagents explore and repair formal proof branches.