Orchestrate a Multi-Agent Research Pipeline
Split a research task across three specialized agents, planner, searcher, and synthesizer, so each one does less and does it more reliably than a single agent trying to do everything at once.
Download workflow config (JSON)How the data moves
This is a handoff plan, not an automatic integration. Review outputs before the next step. Unlabeled arrows show order, not a verified data format. Use the capability links to explore tool substitutions.
Step 1
Split the Job Into Three Roles
Define a Planner (breaks the research question into 3-5 sub-questions), a Searcher (retrieves grounded sources for each sub-question), and a Synthesizer (writes the final brief from what the Searcher found). Write down exactly what each role hands to the next: this handoff contract is what keeps the pipeline from collapsing into one confused agent.
Tools
No catalog tool specified
Step 2
Build the Crew
Use CrewAI to define the three agents with their roles, goals, and the order they hand off work in. CrewAI's role-based design maps directly onto the Planner/Searcher/Synthesizer split, so you're configuring the pattern rather than fighting the framework to express it.
Step 3
Give the Searcher Real Retrieval, Not Just a Model's Memory
Wire the Searcher agent to Perplexity's API for live, cited web search instead of letting it answer from the model's training data. This is the single most common failure point in agent research pipelines: an ungrounded Searcher just hallucinates sources with confidence.
Step 4
Add a Verification Pass Before the Final Brief
Route the Synthesizer's draft through a separate Claude call whose only job is to check that every claim traces back to a source the Searcher actually retrieved, flagging anything that doesn't.
Step 5
Know Where This Still Needs a Human
This pipeline is reliable at gathering and organizing sources, it is not reliable at judging whether the Planner scoped the original question well. Skim the sub-questions before the Searcher runs, especially on the first few runs for a new topic: a poorly scoped plan produces a confidently wrong brief just as fast as a well-scoped one produces a good one.
Tools
No catalog tool specified
Time Required
3-4 hours, one-time setup
Expected Result
A working three-agent pipeline that takes a research question, plans sub-questions, retrieves grounded sources, and hands back a synthesized brief with citations, instead of one agent guessing across the whole task at once.
Maintenance Reality
Every multi-tool workflow needs upkeep. Here's what this one typically needs.
- Typical upkeep
- 2-4 hrs/month
- Single point of failure
- The Searcher's retrieval API: if it goes down or gets rate-limited, the whole pipeline stalls since the Synthesizer has nothing grounded to write from.
What typically breaks
- • The Searcher's retrieved sources go stale, or the search API changes its response shape after a provider update, silently breaking the hand-off to the Synthesizer.
- • Planner scoping drifts over time as the underlying model updates, producing sub-questions that are technically answered but miss the point of the original research question.
Handover tip: Keep the three agents' role/goal prompts and the exact handoff contract (what each one passes to the next) in one shared doc next to the code, not just in the framework config, so a successor can see the intended design instead of reverse-engineering it from agent behavior.
Related Workflows
Build a Multi-Agent Research Assistant
Chain together multiple AI models, each with a specialized role, to conduct deep research on any topic, fact-check findings, and produce a structured report automatically.
View workflowBuild a Price and Competitor Monitoring System
Stay automatically updated on competitor pricing, product changes, and positioning shifts, without manual checking or paying for expensive intelligence tools.
View workflowBuild a Multi-Agent Research Pipeline with CrewAI
Set up a CrewAI pipeline where specialized agents handle different research tasks in parallel, one searches papers, one synthesizes findings, one checks contradictions, delivering a comprehensive brief automatically.
View workflow