Multi-Agent Collaboration

Definition

Multi-agent collaboration is the use of multiple specialized or peer AI agents to divide work, exchange context, review outputs, explore alternatives, and coordinate toward a shared task.

Current Synthesis

Across the bounded sources, multi-agent systems are valuable when decomposition, parallelism, role specialization, or independent checking creates more benefit than a stronger single-model pass. Their reliability does not come from agent count: it comes from an Agent Harness that defines tasks, roles, communication, permissions, shared state, acceptance criteria, and external verification. The newest introductory source reinforces this boundary by showing how rapid delegation can distort original intent and amplify errors before a human notices.

Key Claims

• Multi-agent work can improve review, parallel exploration, handoffs, and recovery from one agent’s context drift.
• More agents do not inherently improve correctness; coordinated agents can agree on an ungrounded result or amplify an early mistake.
• Reliable collaboration needs explicit task definitions, role boundaries, communication protocols, permissions, and externally inspectable success criteria.
• A main-agent or manager pattern can route tasks and evaluate results, but it adds another control point that also needs verification.
• Shared workspaces require task claiming, identity refresh, memory boundaries, and resource isolation to prevent duplicate or conflicting action.
• Multi-agent value is workload-dependent because collaboration can multiply token cost, latency, review burden, and operational complexity.
• Human judgment remains responsible for deciding whether cross-checking, a stronger single model, deterministic tests, or no autonomous action is the right design.

Evidence

• Verification and drift: 贾扬清：我所经历的「人工智能已死」到「AI 颠覆世界」的数年巨变丨串台「声东击西」S10E24, E242｜最快半年AI跑通自进化？与陈天桥首席科学家聊聊硅谷模型必争之地, and EP 20: Understanding AI Agents: From Basics to Future Potential make external checking and intent preservation central.
• Context exchange and parallel work: 当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent and 138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权 describe cross-checking and parallel research loops inside harnessed systems.
• Interface and orchestration: 268. AI时代，个人工作台会重新回到手机吗？ shows answer comparison and a possible main-agent routing layer on a mobile workbench.
• Organization-scale coordination: 用 Agent 动力学，和 40 个 Agents 一起为「人 + AI」做产品｜对谈 Slock.ai 创始人 RC and 我们是如何定义 OpenClaw for Teams 新产品形态的｜对谈 Kuse&Junior 联创兼 CTO 宇豪 ground task claiming, identity, culture, memory, machine separation, and enterprise permissions.
• Cost boundary: E249｜Token经济转点：OpenClaw、Hermes到本地自研的Agent进化之路 treats multi-agent review as potentially useful but materially more token-intensive than a single pass.

Counterevidence & Qualifications

The sources are mostly practitioner interviews and product narratives rather than controlled comparisons. They show plausible mechanisms and operating problems, but do not establish a general reliability gain, optimal agent count, or cost threshold. A stronger model, deterministic verifier, conventional workflow, or human team may outperform an agent debate. Multi-agent agreement is not independent evidence when agents share models, prompts, data, or failure modes.

What Changed

• Migrated the page to synthesis-v1 and compressed source-led additions into claim-grouped evidence.
• Added the introductory distinction between specialized compound-AI systems and action-oriented single agents.
• Strengthened the intent-drift warning: fast agent-to-agent invocation can amplify errors before review.
• Made workload fit and marginal review gain, rather than agent count, the decision boundary.

Related Concepts

• Agent Harness - supplies orchestration, context, permissions, tools, and evaluation boundaries.
• Subagent Workflow - delegation pattern that may use specialized background agents under a primary agent.
• AI Verification - external evidence and tests needed because agent agreement is not proof.
• Agent Task Claiming - coordination mechanism that prevents duplicate work in shared channels.
• Persistent Agent Memory - durable context that can support or contaminate collaboration depending on its boundaries.
• Human Judgment Under AI - accountability layer for goals, acceptance criteria, and consequential action.