Multi-Agent
Multi-Agent System
What Is a Multi-Agent System?
- Multiple agents collaborate in defined roles
- A team beats a lone all-rounder
- Coordinate with orchestration
The first thing that changes when you adopt a multi-agent setup is not the quality of the output but the unit in which work is divided. When a single prompt handles everything from research to writing, you can’t tell where things went wrong when results disappoint, so you end up rewriting the instructions over and over. Splitting roles leaves an intermediate output at each stage. Whether the sources were thin, the draft was off-target or the review was lax, you can point to the stage that failed and fix only that stage. Manageability improves before accuracy does.
There are three basic structures. In the supervisor model, an orchestrator assigns the work and merges the results. In the sequential model, one agent’s output passes straight to the next. In the cross-checking model, two or more agents solve the same problem independently and compare answers. Most Korean companies running agents in practice use a compromise: a supervisor model with a sequential flow layered on top. It comes at a price, though. Each added agent multiplies model calls and tokens rather than simply adding to them, and wait time stacks up with every step. For a job with three or four steps and a couple of tools, a single agent is faster and cheaper.
The line between this and an agentic workflow is also easy to blur. In a workflow, a single agent makes a plan, alternates between tools and iterates on its own; in a multi-agent setup, several actors each have their own instructions, permissions and tool lists. The test for when to split is surprisingly simple: when conflicting rules start to creep into a single instruction. Put 'write concisely' and 'footnote every source' in the same prompt and the model strikes a different compromise each time. The same goes for when permission to read the internal database and permission to search the external web need to be kept apart.
The most common misconception is that adding more agents makes the system smarter. In practice, if an earlier stage copies a single number wrong, later stages accept it as fact and polish it into a plausible sentence. A review agent that never reopens the source won’t catch this error either. Without a stop condition, two agents can trade revision requests until calls pile up by the dozens, and some teams only find out from the end-of-month bill. Before you start, document each agent’s input and output format, its retry limit and the points where a person steps in to break the loop.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayRooted in research on distributed AI and multi-agent systems, it has drawn attention as a practical collaboration architecture alongside the rise of LLM agents.
Consider a hypothetical case. Company A, a public-sector SI firm with 180 employees, had a five-person proposal team handling an average of 12 bids a month. With more than 200 requirements per RFP, missed requirements got them disqualified three times last year alone, and working weekends in deadline week had become routine.
- Proposal team lead Mr. Kim set aside the month of August and split the work across four agents: requirement extraction, similar-case search, draft writing and requirement cross-checking. His reasoning: 'When one agent does everything, the requirements table is wrong from the start.'
- The requirement-extraction agent split the RFP source PDF clause by clause into a table of 213 items, and was told to tag each item with its page number in the original. Those page numbers became the only basis for verification in later stages.
- The draft-writing agent pulled similar passages from 47 proposals won over the past three years and filled in a draft for each section of the outline. Ms. Lee from project management said, 'A draft that took two days now takes half a day.'
- The cross-checking agent matched the draft against the requirements table and flagged 18 unmet items in red. Mr. Kim opened the cited source pages himself, confirmed 14 and sent 4 back as false positives.
- Three months later, hours spent per proposal fell from 62 to 34, disqualifications from missed requirements dropped to zero, and the same five people were handling up to 17 proposals a month.