Practical multi-agent orchestration in Codex

📝 摘要 #

本文介绍了 OpenAI Codex 平台上的多智能体编排实践，以 GPT-5.6 Sol 模型为核心，展示如何通过 Multi-Agent V2 工具实现智能体间的任务委托、信息共享与协调执行。文章提出了一套按推理努力档位匹配角色分工的实用模型：Scout（Sol Light）负责窄范围只读探索，Worker（Sol Medium）处理具体实现，Smart Worker（Sol High）承担困难任务与协调。

核心概念： 多智能体编排的关键在于将推理努力与任务难度匹配，并通过协调者（Coordinator）统一调度、避免重复、保持并行效率，而非让所有智能体都开最大推理。

关键技术栈 / 关键要点：

1. 推理努力分层（Scout/Worker/Smart Worker） — Sol Light/Medium/High 分别对应探索、实现、攻坚三种角色，降低低优先级任务的推理消耗
2. 协调者模式 — Coordinator 作为主调度器分配任务、避免重复调查、跟踪各 Agent 状态；并发默认 4 个 Agent
3. 直接消息传递 — Agent 通过公共消息系统与独立收件箱互相通信，Scout 发现依赖信息可直接传给 Worker，无需 Coordinator 中转
4. 上下文继承策略（fork_turns） — 继承父对话历史让 Agent 理解全局目标；fork_turns: "none" 给 Agent 干净聚焦的上下文，适合只读探索任务
5. 叶 Agent 边界约束 — 给不应再委托的 Agent 明确边界指令："直接完成此任务，不要生成其他 Agent"
6. Skill 模式复用 — 将编排策略封装为可复用的 Skill，给 Coordinator 常驻指令：并行派 Scout、Medium 处理常规实现、High 处理难题、结果汇总、审批留给用户

关键结论 / 实践建议： 多智能体编排的核心旋钮是推理努力、上下文继承、委托权限和协作方式——先从默认配置开始，再逐步实验调优，目标是让团队高效推进工作而不浪费推理算力。

GPT-5.6 Sol gets especially interesting when it has a team to work with. Codex's new Multi-Agent V2 tools give Sol and Terra a natural way to delegate tasks, share updates, and coordinate through complex tasks.

Ultra makes agent coordination the default and is best reserved for high-stakes work, where ambiguity or scattered context justify the added depth.

For other tasks, a short prompt or skill can encourage the same collaborative behavior from Sol Medium, as it stays in conversation with you while organizing the work behind the scenes. With the right nudge, Sol can turn broad requests into focused assignments, bringing in other agents, and deciding when a problem calls for deeper reasoning.

Match reasoning to the work #

While it's possible to have Sol delegate to another model like Terra, the simplest setup keeps one model family and adjusts only the reasoning effort, with dedicated roles like so:

• Scout — GPT-5.6 Sol Light. Answer narrow, read-only questions: locate files, trace a code path, or find relevant tests.
• Worker — GPT-5.6 Sol Medium. Implement scoped changes, run checks, or handle supporting work.
• Smart worker — GPT-5.6 Sol High. Take on difficult implementation, resolve ambiguity, or coordinate help when useful.

Treat these roles as useful defaults. Sol Light retains the judgment to find useful context without spending as much reasoning on discovery.

Let the team coordinate #

The coordinator acts as the primary delegator: it assigns substantive work, avoids duplicate investigations, and tracks what each agent is doing. Scouts can investigate in parallel, while workers can share implementation when responsibilities are clear.

Agents can also message one another directly through a common messaging system with separate inboxes. When a scout discovers something a worker needs, it can recognize the dependency and pass along its findings without waiting for the coordinator to relay them.

Concurrency is configurable per thread and defaults to four agents, including the coordinator. Within that budget, a smart worker might coordinate a scout and another worker, or the coordinator could send three scouts to investigate separate questions.

Choose what context agents inherit #

Forking conversation history helps agents understand the broader goal and earlier decisions. Starting with fork_turns: "none" gives an agent a fresh, focused assignment instead. Fresh-context agents can still recognize when a teammate needs information and contact them independently.

Agents that inherit their parent's context may also see its orchestration instructions. When an agent should remain a leaf, give it a short boundary:

Complete this assignment directly. Do not spawn other agents; your parent's delegation instructions apply only to your parent.

Fresh-context agents won't inherit task-specific tool or safety boundaries, so include any essential restrictions directly in their assignments.

Capture the pattern in a skill #

A practical skill can give the coordinator a few standing instructions:

Stay available to the user while delegating substantive work. Send focused, read-only scouts out in parallel with reasoning_effort: "low" and fork_turns: "none". Use reasoning_effort: "medium" for routine implementation and reasoning_effort: "high" for harder problems. Give each agent clear ownership, avoid overlapping assignments, and tell leaf workers not to delegate. Bring the results together and keep approvals with the user.

Experiment with the knobs #

Start with these defaults, then experiment with reasoning effort, context inheritance, delegation authority, and how agents collaborate. The goal is to understand which settings help a team move the work forward without spending more reasoning than the task requires.