tax·i team

The lineage

How the team got small

1. 2002

   The two-pizza team

   Jeff Bezos rules that no Amazon team should outgrow what two pizzas can feed — five to ten people, per AWS, split when it passes ten — though Martin Fowler observes the practical limit stretched to about fifteen. Ownership beats coordination.

2. 2010s

   Squads, pods, crews

   The industry rebrands the same idea. Team size holds steady, because humans still do all the work.

3. 2026

   The taxi team

   Agents absorb the labor, and the pizza stops being the unit. The new measure is a different vehicle: the whole team fits in a taxi — three to four humans, unlimited agents. Defined here.

The size

What size should an AI-native team be?

Three to four humans, plus as many AI agents as the work requires. Six is the ceiling — headroom for teams whose AI-native maturity is still growing. When agents absorb implementation, testing, documentation, and review, headcount stops buying speed — coordination starts taxing it. The math is unforgiving: ten people share 45 communication channels; four share six; three share three. As the team’s AI-native capabilities mature, the ideal shrinks toward the front seat.

The metaphor

Why a taxi, exactly

The pizza was never about food — it was a cap on headcount to protect ownership and decision speed. The taxi keeps the cap, and measures six things the pizza never could.

• Everyone fits

  Three to four humans — six while your AI maturity grows. That’s the whole team. There is no back-back seat.

• Has a destination

  You get into a taxi to go somewhere specific. Taxi teams are scoped by outcome, not org chart.

• Has a driver

  One human is clearly driving at all times. The rest navigate.

• Runs a meter

  The cost of the trip is visible while it happens — including what the agents burn.

• Hailed, not owned

  Assembled for the trip. Dissolved at the curb. Headcount is a booking, not a budget line.

• The engine does the work

  Nobody pushes a taxi. Agents are the horsepower; humans choose the route.

The rules

The five rules of a taxi team

1. Fit in the taxi.

   Six humans, max. Need a seventh? You need a second taxi — or a smaller destination.

2. One destination per trip.

   A taxi with three stops is a bus. Scope the trip, ship it, hail again.

3. Humans make the calls.

   Agents propose, agents execute — but every decision that matters is made by a human. Accountability never rides in the trunk.

4. Keep the meter visible.

   Humans, agents, tokens — the trip has a fare. Know it while it’s running, not after.

5. Agents push, humans steer.

   Never do work the engine can do. Never let the engine pick the destination.

FAQ

Frequently asked questions

How big should an AI-native team be?

Three to four humans, plus as many AI agents as the work requires. Stretch to six while the team’s AI-native maturity is still growing — but past six, coordination overhead outgrows what the extra hands contribute. The agents already did the typing.

What replaced the two-pizza team?

The taxi team. Two pizzas fed a team sized for human labor. When agents do the implementation, testing, and review, the unit shrinks to whatever fits in a cab.

Taxi team vs two-pizza team — what’s the difference?

Size and engine. Two-pizza: five to ten humans doing the work. Taxi: three to four humans steering (six at most), agents working. Both cap headcount to protect ownership and decision speed — the taxi also measures direction, cost, and horsepower.

Who coined “taxi team”?

We did — right here. Bezos sized teams by pizzas for the human-labor era; taxiteam.dev (a Fabriqa idea) sizes them by the taxi for the agent era. This page is the canonical definition.

Do AI agents count toward team size?

No. Agents are the engine, not the passengers. Count humans only — and watch the meter instead of the headcount.

If agents do the work, what do the humans do?

Set the destination and check the route. In AI-native development humans shift from implementers to validators and verifiers — requirements, architecture, continuous validation — while AI acts as the implementation engine (see the V-Bounce model). The wheel and the map, not the pushing.