AI Agents in Group Chat: Octo Solves Multi-Agent Coordination
Mininglamp Technology's Octo open-source platform reimagines AI agent orchestration by making agents first-class participants in group chat conversations, solving the session island coordination crisis.
The Coordination Crisis in Multi-Agent Systems
Most teams building with large language models eventually hit the same wall: a handful of capable agents — one for data extraction, another for summarization, a third for scheduling — yet the only thing connecting them is a human copying and pasting between browser tabs. Individual agent performance has soared, but inter-agent coordination remains an architectural afterthought.
Engineers at Mininglamp Technology encountered this problem repeatedly while deploying AI agents in enterprise environments. "After a while, using agents felt more exhausting than not using them, because there were simply more contexts to keep track of," the team wrote. Their response was Octo, an open-source platform that treats agents as first-class participants in group chat conversations rather than isolated API endpoints.
The Session Island Problem
Every agent framework optimizes for a single metric: how well one agent performs in isolation. Bigger context windows, faster inference — all valuable, but they miss a structural limitation visible only in multi-person, multi-step workflows. Mininglamp calls this the "session island problem": an agent's context is bounded by its own conversation thread, with no awareness of what another agent decided or flagged upstream. As the number of agents grows, this coordination tax scales linearly.
In a typical enterprise scenario, a marketing team member asks a pricing question requiring product specs from engineering, cost models from finance, and competitive intel from sales. Someone must manually route information between agents. The agents are powerful; the connective tissue is duct tape.
Why an API Gateway Falls Short
The team's first instinct was a conventional API gateway. Three problems surfaced. State management: API calls are stateless, but agent collaboration is long-running, spanning hours with human approvals and intermediate artifacts. Observability: Debugging an API-based agent mesh means chasing logs across half a dozen services; an IM-based architecture gives a single chronological audit trail. Permission management: API gateways use tokens and ACLs; group chat maps visibility directly onto group membership — no separate matrix to maintain.
Octo: Agents as Chat Participants
Octo's core design principle is that agents — called Lobsters — are not invoked through webhooks. They are participants in conversations with the same capabilities as human users. The platform runs on octo-server, a Go backend handling REST APIs, WebSocket connections, and IM routing through WuKongIM.
When a Lobster joins a group, it receives the full conversation context — chat history, roster, read receipts — not just a trigger event. It can proactively message, reply, notify teammates on completion, or ask follow-ups. IM protocols enforce strict message ordering, meaning every agent sees the same history in the same sequence — far more reliable than the request-response pattern of an API mesh.
Scale Out, Not Just Up
Improving a single agent is "scaling up." What enterprises need is "scaling out" — organizing specialized agents to collaborate. Octo's groups let you add agents and humans to project channels where they interact through the same messaging interface. A companion service, octo-smart-summary, uses LLMs to periodically compress chat history into key decisions and action items.
Security and Engineering
Octo ties permissions directly to group membership: when an agent joins a channel, it inherits that channel's ACL. There is no separate permission matrix for agents. For edge-agent scenarios running on local devices, Octo's adapter bridges execution results securely into IM groups without exposing raw data.
Real engineering challenges remain: unifying message formats for structured agent output, handling concurrency conflicts when multiple agents reply simultaneously, and compressing indefinitely growing group histories for finite agent context windows. Octo addresses these with a unified message schema, server-level conflict detection, and LLM-powered summarization.
Open Source and the Road Ahead
The entire Octo project is released under Apache 2.0 on GitHub, including backend, web client, mobile apps, and admin console. The team adopted a local-first philosophy — chat records, vector indices, and agent execution can run on the user's own infrastructure. As organizations move from deploying individual agents to orchestrating multi-agent workflows, the future of enterprise AI may look less like microservice meshes and more like the group chats we already use every day.
