Field Guide · teams
Multi-Agent Teams: Why Three Specialists Beat One Generalist
A specialist agent does one job all the way down. A generalist agent does several jobs poorly. The math on three specialists working in parallel beats one generalist working in series, and Slack is what makes the parallel readable.
The marketing for AI agents has a recurring temptation: ship the all-purpose agent that handles everything. The version-one-of-everything pitch sells well in a demo because it answers every question with yes. It does not survive the first quarter of production. Real teams ship work the way real teams scope work — by role. The agent that handles competitor monitoring is not the same agent that handles outbound drafts; the agent that runs CS triage is not the same agent that handles CS knowledge-base maintenance. This piece is about why the right shape is three specialist agents working in parallel, not one generalist agent doing more, and how the parallel becomes legible to the team in Slack.
Why it matters
A generalist agent forces a buyer into the same mistake teams have been making with AI for two years: the agent gets one constitution that has to cover every kind of work, one corpus that has to span every domain the agent might touch, one set of tool permissions that has to authorize the broadest case. The constitution becomes vague. The corpus becomes shallow. The permissions become permissive. Specialist agents — three or four narrow agents instead of one wide one — invert all three: each constitution is sharper, each corpus is deeper, each permission set is tighter. The team gets more work done, and the team trusts the work more.
Take the marketing function as a worked example. A team that wants AI in marketing has at least three jobs that AI could plausibly do.
Inbound monitoring.
Watching what competitors announce, what customers say in public, what the metrics on the marketing dashboard signal. The agent's stream is alerts and feeds. The agent's output is a daily roll-up in Slack with the questions worth thinking about. The Roster's VYRA-01 — Mira, the inbound BDR agent — runs a version of this for the lead-detection slice.
Drafting and brief production.
Taking a directive — a launch brief, a campaign concept, a competitor response — and producing the structured first cut. The agent's stream is requests from the team. The agent's output is a draft in the doc, with a Slack post linking to it. VEXA-01 — Beatrix, the marketing strategist — runs the brief-production version of this on the Roster.
Knowledge stewardship.
Keeping the marketing wiki current, the brand guidelines accessible, the prior briefs searchable. The agent's stream is content-management events and the team's questions. The agent's output is search-ready answers and a wiki that does not rot.
A team that tries to ship one agent to do all three runs into the obvious problem: the constitution that authorizes brief production conflicts with the constitution that limits autonomous wiki edits, the corpus needed for monitoring is different from the corpus needed for drafting, and the permissions for the monitoring stream are different from the permissions for the brand voice. The single-agent version compromises on each axis. The three-agent version does not.
The economic shape: half an AI per human role.
The math the platform commits to is roughly half an AI per human role. Three agents handle most of the daily output of one role; the human in that role does the work that didn't go to the agents. Across the marketing function, that means three specialist agents to start, the human marketing lead steering them, and the team's daily output increases roughly in proportion to the work the agents handle. Compounding the load with one generalist agent does not produce the same lift; the human marketing lead spends the time saved on covering the agent's mistakes.
Slack is what makes the parallel work.
The reason three agents in parallel are not three times the noise is that all three post in the same set of channels, with conventions about which channel each posts in and which thread their work goes to. The team scrolls one feed. The work is differentiated by author — VYRA-01's monitoring posts look different from VEXA-01's brief drafts — and converges into a single readable timeline. Move the same three agents into three separate dashboards and the team has to remember to open three places. Keep them in Slack and the team reads them the way they read any three coworkers.
What this is not.
A multi-agent team is not the same as an “agent swarm.” A swarm is many agents acting together on the same task; a multi-agent team is several agents acting on different tasks that share an audience. The first is a research direction. The second is a deployment pattern that already works.
The edge
Specialization is what most production AI gets wrong, because demos reward generality. A demo-stage buyer wants to see one agent do four impressive things in five minutes. A production-stage buyer wants to see three agents do one task each, very well, every Monday for the next year. The economics favor the second. The hiring shape — three specialists at half the AI-to-human ratio — produces more output per dollar than one generalist at the same total spend. The reason vendors pitch the generalist anyway is that the generalist closes the deal faster. The reason the team should refuse the generalist is that the generalist breaks first.
Honest take
Three agents requires three constitutions, three onboarding processes, three sets of channel permissions, three escalation paths. The honest cost of multi-agent teams is integration overhead at the start. The first month with three agents is more work than the first month with one. The productivity inverts by month three, and stays inverted, but the buyer who has not been told that the first month is heavier sometimes confuses the integration cost with the agent's real cost. The platform's job is to keep that integration cost low — templated agents, standard channel patterns, shared escalation conventions — but it does not become zero. The buyer should know.
Three specialist agents working in parallel, each doing one job all the way down, posting into the same Slack the team already reads. That is the shape that scales. The all-purpose agent is the shape that closes the demo and breaks in production.