Podcast Atlas
concept Updated 2026-07-08 Tags: Agents, Workflow, Skills

Subagent Workflow

Subagent workflow is an agentic pattern where a foreground assistant delegates complex, long-running, or adversarial tasks to background agents and later integrates their outputs. In 阿里千问离职余震，在几万人的铁球里如何体面生存, the host describes one skill that sends tool-heavy or high-token tasks to a background subagent and another that has pro and con agents debate a question before synthesis.

探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐 adds the governance side of subagent work. Lai Xinlu argues that multi-agent systems need role-specific permissions, information boundaries, and handoff documents, such as giving a code-exploration agent read-only powers while preventing a test agent from changing production code just to make tests pass.

当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent adds the cross-checking side through MiniMax. The source says multiple agents can exchange much larger context than human feedback usually provides and can correct a single agent when long-context work starts drifting.

Vol. 166 闲聊: 从 Gemini 到 AI 的加速与混沌 adds a personal coding workflow side: subagents let Superpowers, Claude Code, and Codex continue long tasks without filling the main conversation, while review loops let agents check and repair each other’s output at the cost of more tokens and more human supervision.

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 adds an AI For Math case. Hong Letong / 洪乐潼 says Axiom used subagents, experience learning, and skills as alternatives to costly broad search, suggesting that theorem proving can use specialized agents for proof planning, tactic attempts, formalization, and verifier-driven repair.

Key Patterns

  • Background execution for tasks too large or disruptive for the foreground thread.
  • Reference IDs for later retrieval, pausing, or management.
  • Adversarial analysis by assigning different agents opposing roles before a synthesizer reviews them.
  • Role-specific tool permissions and information access.
  • Context compression and handoff documents when a new agent continues work from a prior agent.
  • Peer review or cross-checking between agents when a long task needs another context window or viewpoint.
  • Main-context preservation in practical coding workflows where the user wants several tasks moving while still keeping one controlling thread clear.
  • Theorem-proving workflows can use subagents to explore proof branches, repair Lean failures, and preserve successful tactics for later reuse.

Connections