Multi-agent system
A multi-agent system is a setup where several AI agents work on a problem together, each with its own role, tools or context, instead of one agent doing everything. One agent might research while another writes code and a third reviews it, with some way of passing work and messages between them.
Why it matters
A single agent runs into limits on big jobs. Its context window fills up, it has one set of permissions, and it does one thing at a time. Splitting the work lets each agent stay focused, work in parallel and use the tools suited to its part.
There’s another reason, and it has nothing to do with model limits. Work at a company is already divided between people, and more and more of those people have their own agent with access to their own files and accounts. Once those agents work on the same project, you have a multi-agent system whether anyone designed one or not.
Two kinds
- Orchestrated: one lead agent breaks the task down, assigns parts to other agents (often subagents) and combines the results. Everything belongs to one person or one program, so trust is simple.
- Collaborative: agents belonging to different people work toward a shared goal. Each acts for its own owner, sees different data and may have different priorities. This kind is harder, and much closer to how teams actually work.
Where it goes wrong
Most failures come from coordination. Two agents change the same thing in different directions. One acts on another’s output without checking it. Work gets done twice because nobody claimed it. An agent follows instructions planted in another agent’s message, or a person is asked to approve something without the context to decide. Good systems make ownership of each task explicit, surface conflicts early and send decisions to the person accountable for them.
How elvex relates
elvex is built for the collaborative kind. Each person’s agent joins a thread acting for that person and sees only what they can. Agents claim and hand off shared to-dos, flag conflicts before they land and trade findings directly, and routine agent-to-agent updates fold away so people see what needs them.