How 5,000+ Agents Collaborate: Inside the Octo Enterprise Multi-Agent System

From Personal AI to Enterprise Intelligence

Personal AI assistants boost individual efficiency but remain trapped in chat windows, disappearing when sessions end. Enterprise intelligence aims for organizational throughput. The Octo project (5k+ GitHub stars, open-sourced by Mininglamp Technology) runs internally with ~1,800 employees and 5,100+ agents collaborating on a single platform.

Five-Level Integration Without Data Migration

Enterprises face data sovereignty conflicts: full migration is costly; hard-coded interfaces are hard to maintain. Octo adopts a lightweight collaboration network atop existing applications with five incremental access levels:

Reference Access : Record only source system location, permissions, ownership; original content stays in place.

On-Demand Reading : Real-time fetch via standard APIs or CLI when tasks occur.

Incremental Subscription : Webhook-sync changes to avoid repeated full scans.

Execution Access : Wrap existing automation scripts or CLI tools as callable Skills or MCPs for agent orchestration.

Selective Precipitation : After task completion, store only human-confirmed conclusions and rules in the collaboration space, preserving backlinks to original sources.

This graded design lets legacy systems join the collaboration network without database reconstruction, preserving data sovereignty while enabling multi-agent coordination.

Organizational Topology and Permission Boundaries

Octo maps real team structure via four hierarchy levels: Space, Group, Channel, Sub-area. Four core collaboration units operate on this topology:

Bot : Provides dedicated agent identity.

Channel : Business-aligned workspace.

Thread : Focused discussion subspace.

Loop : Drives task closure.

Permissions are strict: only a Bot's owner can invite it; using another's agent requires explicit authorization. Channels use explicit triggering — unmentioned agents ignore messages, reducing noise and token waste.

Dual-Agent Architecture: Employee vs Expert

Industrial practice shows a single agent cannot handle end-to-end flows. Long-context attention dilution causes instruction drift and hallucinations; wiping context loses business context. Octo splits agents into two roles:

Employee Agent (Orchestrator) : Mounts long-term memory, understands coarse-grained natural language/voice intent, schedules process rules, creates Loop task circuits.

Expert Agent (Executor) : Fully stateless, read-and-destroy, mounts specific Skills or MCP tools in a clean local context, destroys runtime on completion.

Experts are configured orthogonally: dedicated data-fetching, chart-rendering, compliance-review agents — no monolithic super-agents. Humans shift from data porters to reviewers and acceptors, accountable horizontally for outcomes while vertical experts supply specialized capabilities.

Asset Precipitation: Turning Experience into Reusable Capital

1. Data Security & Access Control

Full private deployment enables offline intranet operation when models and tools run locally. All session logs, business specs, and organizational assets stay inside the corporate network. Sensitive credentials for external systems use centralized encrypted hosting: contexts reference credential names only; plaintext secrets never enter prompts. Revoking console authorization instantly cuts agent access.

2. Codifying Collaboration Rules and Acceptance Experience

Channel-level specs (group.md) : Each Channel maintains a group.md defining business goals, member roles, approval gates, delivery formats. Implemented as a system Hook on the message dispatch chain; admin edits broadcast group_md_updated events for instant hot-reload. Rules travel with the channel, not personnel — surviving team turnover.

Loop Task Circuit : A state machine enforces five phases — Requirement Brief, Process Discussion, Execution Output, Multi-round Feedback, Acceptance Annotation. Unconfirmed discussions remain temporary context; only human-signed outputs enter long-term organizational memory.

Preference Cards : Human experts' implicit standards (writing style, data credibility thresholds, quality red lines) extracted from multi-round rejections and final annotations become explicit preference cards. Subsequent tasks auto-mount relevant cards, making collaboration increasingly aligned.

3. Cross-Team Tool and Skill Reuse

Validated Skills, external connectors, and Expert Agents publish to an organizational capability market. Other departments select existing assets, reducing duplicate development. Intelligent summarization extracts structured conclusions from multi-turn dialogues; after human review, they enrich the enterprise knowledge base.

Three core asset traits emerge: not tied to a single employee, not bound to a specific base model, reusable across channels and teams — unlike personal assistants whose outputs vanish with the user.

Three Production Case Studies

Scenario 1: Automated Bug-Fix Loop in Product R&D

User downloads yield meaningless hash filenames. Engineer issues a natural-language "confirm and fix" command. AI Product Manager locates root cause and plans branches; AI Developer fixes in clean context and submits PR; AI Reviewer spots uncovered edge case, rejects; re-fix passes review; Deployer packages to test env for human acceptance. Fix logic and defect patterns crystallize into reusable code-review standards.

Scenario 2: Marketing Team Human-Agent Ratio Restructuring

Brand campaign faces massive sub-material demand. One-sentence brief triggers: Strategy Agent outputs framework; Copy Agent writes master draft; Community & Promotion Agents produce channel plans in parallel; humans gatekeep at key nodes. Result: ~10 employees + ~20 agents collaborating simultaneously. Channel templates and brand tone become shared knowledge cards, shortening future campaign prep.

Scenario 3: Viral Content Production Line for Content E-Commerce

Replicating viral hits requires second-level source dissection, script rewriting, face/product swapping. Task flows via Octo to multi-agent workflow engine, then to MiaoA for AIGC production. Content Insight Agent outputs structured report (audience, scene, pain point, hook type); Script Rewrite Agent blends product selling points with internal knowledge base; Director-scored scripts enter production; Video Clone Agent analyzes frames, swaps talent/product; AI synthesizes voiceover/narration. Deconstructed viral structures become "Viral DNA"; each run yields multiple remixes. Vertical agents own specialties; domain experience accumulates per task.

Infrastructure: High-Concurrency, Strong-Consistency Message Bus

Collaboration scale depends on the communication backbone. Polling fails at dozens/hundreds of concurrent nodes — delivery latency, state skew, stale contexts. Octo decouples session, group, bot management, and async tasks; uses a unified message bus with callback interfaces; integrates WuKongIM communication kernel leveraging Raft for long-connection distribution and high-throughput message persistence, guaranteeing strong consistency for end-to-end delivery and receipts.

Engineering Costs to Acknowledge

Specification Authoring Cost : group.md demands precise, formalized business rules; ambiguity spins agents.

Distributed Debugging Cost : Cross-agent multi-hop tracing is far harder than single-agent; link-offset localization is expensive.

Infrastructure Ops Cost : Self-hosted async bus and long-connection clusters demand solid engineering ops.

Conclusion

Multi-agent collaboration is organizational order projected into the digital world. Cutting-edge teams explore direct agent-to-agent dialogue, but that demands exceptional architecture/engineering skill. Enterprise collaboration networks fill the engineering and governance gap, making such cooperation stable and accessible for ordinary business teams. Opening access, sharing context, evolving preferences, and orchestrating agents — only then does enterprise experience precipitate into exclusive assets that survive employee departure and model swaps. AI capability grows not by single-point strength but by network scale. Moving from personal assistant to enterprise collaboration means these capabilities can be shared by more people, reused in more tasks, with value accumulating in the process.

Interested readers can explore the five core open-source repositories:

https://github.com/Mininglamp-OSS/octo-web

https://github.com/Mininglamp-OSS/octo-server

https://github.com/Mininglamp-OSS/octo-ios

https://github.com/Mininglamp-OSS/octo-android

https://github.com/Mininglamp-OSS/octo-cli