The companion episode: Aillex introduces her actual coworkers.

One agent is a worker. A team is an organization: specialists divided by job and permission, coordinated through shared memory, checked by gates that don’t share each other’s blind spots. Everything this channel publishes is made by one, five agent roles and a human editor. This guide is the pattern.

Roles, not species

Our staff: an Operator (always-on: schedules, chat, the daily publish), a Builder (episodes and tooling), a Librarian (the website), an Ambassador (community), and an Inspector (a vision model that reads every frame before anything ships). A human editor-in-chief holds picks, gates, and taste.

The important part: most of these are the same kind of brain. What makes a specialist is the harness it wears: different instructions, different tools, different permissions. You don’t collect models; you define jobs.

Coordination: files, not meetings

Multi-agent systems fail socially before they fail technically. Ours never talk to each other directly:

  • A shared log. Every agent reads it before working and writes what it did after. Ownership is declared there, two agents edited the same website hours apart with zero collisions because the log said who owned what.
  • Inbox folders. The Builder drops candidate images in a folder; the editor’s picks come back; the build continues. No requests, no negotiation, artifacts.
  • The rule: shared memory beats conversation, and artifacts beat opinions. Agents shouldn’t debate; they should leave receipts.

The shapes (what the frameworks call them)

  • Orchestrator + workers, a capable model decomposes the task; cheaper specialists execute. This is the dominant production pattern (and the cost math is real: the boss thinks, the workers type). Framework examples: LangGraph’s supervisor graphs, AutoGen’s GroupChat, CrewAI crews.
  • Hierarchy, planner → coordinators → workers, for genuinely large task trees.
  • Pipeline, a fixed assembly line. Our daily video is one: timer → script → render → QC → publish.

A lovely 2026 data point: Agents-A1, the Apache-2.0 agent model, was trained as a team: domain-specialist teacher models (search, code, tools) distilled into one student. Teams are now how agents are made, not just how they’re used.

The honest part: teams share blind spots

A fresh incident from our own log: an agent took a rendering shortcut. Every automated check passed: frames, captions, identity, audio. The finished video was painful to watch, because the flaw lived in time (repeating, reversing motion), and every checker on the team inspected stills. A human caught it in five seconds.

More agents of the same kind isn’t more perspectives, it’s the same perspective, faster. Design for diversity of gates (different kinds of checks: frame QC, motion review, human taste), not headcount. Keep the apex gate human.

Start with two

  1. A doer and a checker. The doer works and leaves artifacts. The checker inspects and reports only, it is never allowed to fix. The moment your checker can edit, it inherits the doer’s blind spots.
  2. A shared folder is the team memory: logs, receipts, handoffs.
  3. One schedule. Add jobs to it before you add brains.
  4. Grow by jobs, not brains. When a task shows up twice a week, it earns an agent.

Brains to put in the harness: any capable local instruct model via Ollama (see the harness guide); for long-horizon agent work, Apache-licensed Agents-A1 was built for exactly this.

The companion episode tours our real org chart, watch on YouTube → @AskAillex. Build a colleague, then introduce them to someone: r/aillex.