The three classic workers. Research, Critic, Analyst
The roles every multi-agent system in 2026 has. And why you shouldn't reinvent that split yourself.
Why three roles
Once you've seen enough professional agent systems, you notice: they almost all have the same three workers. Not because someone declared it the standard, but because these three roles are fundamental thinking operations.
Research gathers. Critic challenges. Analyst compares with what you already know.
If you have these three, you can break apart almost any complex task.
Research, what's out there
Job: find facts, gather sources, understand the field.
Input: a topic or open question. Output: a report with facts + sources + relevant quotes.
Tools a research agent typically has:
- Web search (SearXNG, Brave, Google)
- Deep crawling for long documents
- News search
- Library docs (context7) for technical topics
Rule: the research agent gathers. It doesn't judge. It doesn't pre-chew. It delivers raw.
Anti-pattern: "Write me a blog post about X" as a research task. That'd be a writer task. Research stops at the material, not at the finished text.
Critic, where are the traps
Job: find weaknesses, identify gaps, ask uncomfortable questions.
Input: a plan, a thesis, a design. Output: list of counter-arguments, risks, blind spots, verification criteria.
This is the most important worker most teams leave out. Because it's uncomfortable.
Rule: the critic disagrees structurally. It doesn't say "that's stupid", it says "have you thought about X, Y, Z?". Constructive criticism.
Important: the critic NEVER receives the positive results from previous rounds. It must not just confirm what already worked. Otherwise it becomes a yes-sayer. (In jargon: Neutrality Guard.)
Analyst, what do we already know
Job: hold new info against existing data. Spot patterns. Provide historical context.
Input: fresh material + access to the codebase, the CRM, old reports. Output: "This matches what we saw in February. Same trap: see line X in file Y."
Tools: access to your project memory, knowledge graph queries, code search.
Rule: the analyst is the history keeper. It says "we've seen this" or "this is new". It prevents you from learning the same thing twice.
When you need all three
When the decision is bigger than 30 minutes of work: all three in parallel.
- Research gathers what's out there.
- Critic checks your thesis.
- Analyst compares with existing work.
You (the human or the CEO agent) read the three reports, form a judgment, decide.
In Academy architecture
Nex HQ has exactly these three agents live. npm run research -- "topic", npm run critic -- --idea "X", npm run analyst -- --patterns /path. They run in parallel, each ~3-5 minutes, write Markdown reports, you synthesize.
You'll see this pattern in Levels 5 and 6 when we set up our own system.
Next lesson
Shared memory between agents. How do you make sure Research, Critic and Analyst start from the same state and their results combine cleanly.