ByHeartAI
Advanced8 min read

Design a Research or Multi-Agent System

Research systems browse and synthesize with citations. Start with one agent and tools. Fan out extra agents only for independent, parallel reads — then cap loops, isolate tools, and trace every handoff.

Explain like I'm new to AI

The multi-agent lesson warned: extra personalities are extra cost and extra bugs. This is when the job is research (many sources, a brief) and when a team actually pays.

A single agent with search, browse, and a note tool. Parallelize reads with extra tool calls, not extra personalities. Default in 2026 until eval says otherwise.

Multi-agent is for parallel, separable research — not a default architecture. Citations and loop caps are the product.

Job: a sourced brief on a question. SLO: citation coverage, time budget, $ cap, no silent tool storms.

Shape: one agent + search/browse/notes is the default (2026). Multi-agent when sub-questions don't share state and can run in parallel.

Mental model

A journalist with a notebook vs a newsroom: you don't hire five editors to fetch one weather page. You do send two reporters to two cities.

How it works

  • Tools: search, fetch, maybe a code interpreter in a jail. Fetch is indirect injection — no send-mail on that kit.
  • Output: claims with source urls/ids. Verify citations in code when you can.
  • Memory: a research pad (notes tool) rather than stuffing every page in the window (compress/isolate).
  • Multi-agent: orchestrator assigns disjoint queries; synthesizer only sees notes, not god refund tools.
  • Eval: rubrics on groundedness + coverage; humans on a sample. Cost per brief is a first-class metric.
  • Traces: parent run id, child spans per gatherer.

Real-world example

"Competitive memo." One agent, 12 fetches, 8 minutes, $0.40, citations. The "team of 6" version spent $9, disagreed, and cited a URL it never fetched. Eval chose the single agent.

Technical explanation

Fan-out helps read-heavy, decomposable work (Anthropic-style guidance you've seen in the agents category). Tight sequential coding/support work gets worse with more agents.

MCP: each gatherer gets a read-only server. Orchestrator holds the write-to-doc tool.

Don't confuse this with AI search (next): search returns a ranked list; research writes a document.

Common mistakes

Common mistake

A Slack channel of agents with every MCP server installed because "collaboration."

  • No $ or step cap.
  • Synthesizer that can also hit production APIs.

When to use it

  • Open-ended briefs over many sources. Fan-out when eval shows wall-clock or coverage wins.

When NOT to use it

  • Don't multi-agent a FAQ. Don't skip citations because the prose sounds like a report.

Alternatives

  • RAG chatbot over a closed corpus. Human researcher + RAG. Plain search UI.

Quick quiz

Question 1 of 3

Default research architecture in 2026?

Question 2 of 3

When is multi-agent worth it?

Question 3 of 3

True or false: a fluent report without fetchable citations is a successful research agent.

Related concepts

  • MCP + Agents, IDEs, and ProductsMCP is the plug; agents, IDEs, and products are hosts that decide when to call — a product can also publish a server so other hosts reach it.
  • Planning & ReflectionPlanning breaks a goal into sub-tasks before acting, and reflection lets the agent critique and fix its own work — together they make agents far more reliable.
  • Tool, Retrieval, and Agent TracesNested spans for retrieve, each model call, and each tool — a waterfall that shows which stage failed, without dumping secret payloads.
NextDesign an AI Search Engine

Last reviewed: 2026-09-04 · Written by ByHeart AI · Reviewed by ByHeart AI