Multi-Agent Systems for Business: When You Need Them
"Multi-agent" is the most over-sold phrase in AI right now. Most businesses do not need one, and the ones that do are usually told they need something far more complicated than the pattern that actually works. Here is the honest version.
- Published
- Updated
When one agent stops being enough
A single agent with a good harness handles more than people expect. It stops being enough in exactly three situations.
- The work fans out. "Research these forty suppliers" or "process each file in this folder" — one agent doing it serially fills its context with reading and gets slower and dumber as it goes.
- The steps need different skills or permissions. The agent that drafts the contract should not be the one with write access to the CRM.
- The task is longer than one context window. A multi-day project with dozens of sub-tasks needs a coordinator that keeps the plan and workers that do the pieces.
The pattern that works: orchestrator and workers
Nearly every successful multi-agent system we have shipped is the same shape. One orchestrator holds the plan, splits the task, delegates and merges. Workers each receive a narrow brief, a small tool set and return a short result. The orchestrator never reads raw data; the workers never see the whole plan.
| Role | Model tier | Context | Tools |
|---|---|---|---|
| Orchestrator | Most capable, high effort | Plan + worker summaries only | Delegate, merge, ask human |
| Reader / researcher worker | Cheaper, low effort | One document or source | Read, search, summarise |
| Actor worker | Capable, medium effort | One task brief | A few write actions, with approval |
| Reviewer worker | Capable, high effort | One artefact + rubric | Grade, comment |
Tooling has caught up with the pattern. Vendor SDKs ship subagents natively, and managed agent platforms support multi-agent sessions where an agent delegates to copies of itself or to a cheaper worker agent by ID. You rarely need to build the delegation plumbing from scratch any more.
What it costs and what it saves
A multi-agent build runs 1.5–2.5× the cost of a single-agent build on the same task, mostly in evals and observability — you now have to trace failures across hand-offs. Running cost is often lower than a single agent, because readers run on cheap models and the expensive orchestrator sees only summaries.
- Build: $40k – $150k depending on the number of worker types and the systems they touch.
- Run: typically 30–60% cheaper per task than one large agent doing everything, once tuned.
- Timeline: 8–14 weeks including two weeks of shadow mode.
How we build and teach it
We start every multi-agent project by building the single-agent version first and measuring where it breaks. Then we split only along the measured seams. For teams who want to own it, our multi-agent workshop covers orchestration patterns, cross-agent tracing and the eval design that makes hand-offs testable.
Frequently asked questions
Do agents talk to each other freely?
Not in systems that work. Free-form agent-to-agent chat is expensive and hard to debug. Use structured delegation — a brief goes down, a result comes back — with the orchestrator as the only coordinator.
Can workers run in parallel?
Yes, and that is the main speed win. Ten reader workers on ten sources finish in the time one would take.
What is the biggest risk?
Losing the thread across hand-offs. Every brief and result must be logged, and the eval set must include multi-step scenarios, not just single-worker tasks.
