Agent Teams Explained - Headquarters + Team Lead Pattern

An agent team is a coordinated set of AI agents working on the same business goal, with a Team Lead that delegates subtasks to specialist Team Members, collects results, and manages the workflow. The term has emerged across the AI category in the last year - but the working product matters more than the label. This page covers what agent teams actually are in cto AI Business, how the Headquarters / Team Lead / Team Members pattern works, and how to spin one up.

TL;DR

An agent team is one Team Lead + multiple specialists, each with:

• A role (research, draft, review, send)
• A model (Auto, or a specific frontier LLM like Opus 4.8 or GPT-5.5)
• A scoped tool set (which MCP servers it can call)

You manage the team from Headquarters - the command center that shows overview, team activity, approvals, and pause/resume token usage. You chat with the Team Lead; the Lead delegates to Team Members; output and tasks land in the kanban Tasks view.

Free forever (ad-supported, rolling 24h+7d limits) to pilot a team end-to-end. Premium raises the limits and unlocks premium agent features.

What's in the product

Headquarters

The dashboard. Overview of your team, current activity, pending approvals, billing usage. Pause/resume controls let you stop token usage immediately when you need to.

Plan

Every AI business has a Plan - your core business objective, written in plain language. The Team Lead reads it; it drives delegation. Refine the Plan by chatting with the Lead or editing it directly. Shorter, more succinct plans reduce token usage.

Team Lead

The primary chat interface. You talk to the Lead like you'd talk to a manager. The Lead:

• Hires and fires Team Members
• Assigns and evaluates work
• Interacts with MCP integrations on the team's behalf
• Manages settings (including which models each member uses)

Team Members

Specialist agents you hire via the Hire employee button. Each member has:

• A name + role (you pick)
• An assigned model (Auto, Opus 4.8, GPT-5.5, etc.)
• An activity log you can audit
• Permissions inherited from the team (which MCP servers they can call)

Tasks

A kanban view of work in progress. Each task has a status, an assigned member, and any output (files, links) attached. You can manually delete tasks; the Team Lead can re-prioritize.

Approvals

Certain actions require your approval before the team executes - surfaced in the chat with the Lead and at the top of the page. You set the approval threshold; e.g., write actions on customer-facing systems usually require human sign-off.

How agent teams are different from a single agent

Agent team topology: lead agent coordinating four specialistsLead agentResearcher read-onlyDrafter writesReviewer criticSpecialist write-scopedPublisher outputLead delegates subtasks → specialists return results → lead synthesizesA small agent team: one lead coordinating four to six role-specialized agents, each with scoped tool access.

Single agent does everything itself in one context window. Agent team splits the work:

• Each member has fresh context for its subtask (no context-window decay).
• Each member has its own tool permissions (a "researcher" agent has read-only MCP; a "publisher" has write).
• Each member can be on its own model - fast cheap for triage, strong reasoning for nuanced work.

The Team Lead routes between members until the goal is achieved, then synthesizes.

What you can build

Production agent-team patterns we see in 2026:

• Inbox triage + reply. Lead reads incoming email → researcher pulls context from CRM (Linear or Notion MCP) → drafter composes reply → reviewer checks tone → publisher sends.
• Customer support escalation. Lead reads ticket → researcher pulls account history (Supabase MCP) and knowledge base (Notion MCP) → drafter composes resolution → reviewer checks compliance → action agent updates the ticket.
• Marketing campaign authoring. Lead reads brief → researcher pulls campaign data (Webflow + Notion MCP) → drafters produce variants in parallel → reviewer compares → Lead picks the winner → publisher ships.
• Code review on PRs. Lead reads diff → specialists per concern (security, performance, style) → Lead synthesizes review and posts.
• AI receptionist. Lead handles call (via Twilio MCP) → booking specialist for appointments → billing specialist for account look-up → emergency-triage specialist on urgent calls.

How to build one in cto AI Business

The flow:

1. Create the AI business. Sign in at cto.new, click "New business," write a Plan (1–3 sentences on the objective).
2. Talk to the Team Lead. The Lead is ready immediately. Describe what you want the team to handle in more detail.
3. Hire Team Members. Click "Hire employee" → name, role, assigned model. The Lead can also hire on your behalf during chat.
4. Wire MCP integrations. At cto.new → Integrations → MCP, connect Sentry, Vercel, Supabase, Cloudflare Observability, Notion, Neon, Linear, Prisma, Render, Webflow (pre-configured), or any custom MCP server. Enable per business.
5. Set approvals. Decide which actions need human sign-off. Defaults are sane; tighten or loosen via Team Lead chat.
6. Run it. Tasks flow into the kanban as the team works. Watch the Tasks view; intervene via Headquarters when needed.

Free tier covers a small team running low volume - enough to validate the workflow. Premium raises the rolling 24h+7d usage windows and unlocks premium agent features.

Auto model + manual override per Team Member

When you hire a Team Member, you assign a model. Choices:

• Auto - cto's gateway picks per task based on the work shape. Default.
• Pin frontier model - Opus 4.8 for hard reasoning, GPT-5.3 Codex for code-heavy work, Gemini 3.1 Pro for long-context analysis, etc.
• Pin open-frontier model - GLM 5.1, Kimi K2.6, MiniMax M2.7 for cost-sensitive workloads.

A team of mixed Auto + pinned members is common: triage members on Auto for cost; review members pinned to Opus 4.8 for quality.

Agent teams vs workflow automation (Zapier, n8n)

Different category, partial overlap:

• Workflow tools (Zapier, n8n, Make) execute deterministic recipes. Trigger → action → action. Excellent for predictable workflows where you know in advance what should happen.
• Agent teams make decisions. The Team Lead picks specialists based on what the task needs. Handles ambiguity, edge cases, and tasks the author didn't anticipate.

Most production automation in 2026 uses both. Simple Stripe webhook → Slack notify? Workflow-first. Triaging incoming customer questions and routing each to the right specialist? Agent team.

FAQ

What's an "AI business" in cto?

A workspace for an agent team with a Plan, Headquarters, Team Lead, Team Members, MCP integrations, and Tasks. Each AI business is one team focused on one objective.

Can I have multiple AI businesses on one account?

Yes - same account, multiple businesses. Useful when you're piloting different agent-team patterns or running unrelated workflows.

Are agent teams ready for production in 2026?

Yes - cto AI Business has customers running teams in production. Maturity varies by use case; production-grade observability is via the Tasks kanban + per-member activity logs.

How are agent teams different from CrewAI or LangGraph?

CrewAI and LangGraph are open-source frameworks - you write code, you operate it, you bring providers. cto AI Business is a managed product - no code, you operate via Headquarters, providers are in the gateway. CrewAI is right when you want code-first control; cto AI Business is right when you want to ship without operating infrastructure.

Can the Team Lead hire on its own?

Yes - within your approval rules. You can let the Lead hire any member, or require approval before each hire. Default is approval-required.

Do agent teams cost more than a single agent?

Per task, yes - multiple model calls add up. But teams complete tasks single agents can't, and you can route each member to a cost-appropriate model (Auto picks cheap for triage, expensive for reasoning). Net cost is often lower for the kinds of tasks you'd previously have hired a human for.

What MCP integrations are pre-configured?

Sentry, Vercel, Supabase, Cloudflare Observability, Notion, Neon, Linear, Prisma, Render, Webflow. Custom MCP servers (stdio or streamable HTTP) via the same UI at cto.new.