How 3–4 humans
ship like 40

1. Seat the crew

A taxi team is three to four humans covering four jobs. Jobs, not titles — one person can hold two, but no job can be held by zero:

• Destination

  Owns what this trip is for and what "arrived" means. Writes the outcome before anyone prompts anything.

• Route

  Owns architecture and decomposition — cutting the destination into pieces agents can drive independently.

• Meter

  Owns cost visibility: agent spend, cloud spend, and time. Reports the fare while the trip is running, not after.

• Validation

  Everyone. Every human validates agent output. This is the job that replaced implementation — it is never delegated to the engine.

Why so few? The communication math is unforgiving. Every seat you add taxes every seat already in the cab, because channels grow with the square of headcount:

channels = n(n−1)/2

And the evidence says the hands aren't the bottleneck anymore — in randomized field experiments with 4,867 developers, an AI assistant alone raised completed tasks by 26% (Cui et al., Management Science). With full agent fleets, the multiplier is the point.

2. Load the engine

Agents do the implementation, the tests, the docs, and the first pass of review. Humans never do work the engine can do — that's rule five, and it's the one teams break first, usually out of habit.

• Spec before prompt. Every piece of work gets a destination (outcome), constraints, and validation criteria written down before an agent touches it. A prompt without acceptance criteria is a taxi without an address.
• Fleet, not intern. Run agents in parallel on independent pieces — the decomposition from the route job is what makes this possible. One agent at a time is a horse chariot with an engine bolted on.
• Agents propose, humans dispose. Agent output is a proposal until a human validates it. No proposal merges itself.

3. Drive by validation

The deepest change is not speed — it's what the humans are for. The V-Bounce model reworks the classic V-model for AI-native development: implementation collapses from weeks to hours, and human effort moves to the two ends that remain hard — deciding what to build, and verifying what was built.

In practice the loop looks like this, repeated in cycles measured in hours or days, not sprints:

1 → 2 → 3 → 4 → hail again · unvalidated output never reaches the destination

If your calendar still looks like implementation — long heads-down coding blocks, standups about progress — you've bought agents but kept the old team. The taxi team calendar looks like short trips: framing sessions, validation sessions, and a lot of agents running in between.

4. Run the meter

The two-pizza team's costs hid in payroll — invisible per feature, reviewed annually. A taxi team's biggest variable cost is agent compute, and it's meterable per trip. So meter it:

• Fare per trip. Every trip gets a budget — tokens, agent-hours, dollars. Not to minimize spend, but to make it a decision instead of a surprise.
• Visible while running. The meter is read daily, by the whole cab, while the trip is in motion. A fare you discover at month-end is not a meter; it's an invoice.
• Spend where it multiplies. A high fare on a trip that ships is a good fare. The meter exists to kill trips that burn without moving — not to make the team afraid of the engine.

This is the operational gap most teams feel first, and it's the one fabriqa.ai is built for: coordinating fleets of coding agents with the meter visible.

5. Keep trips short

One destination per trip. A taxi with three stops is a bus. Scope the trip so agents can finish the implementation inside hours or days, validate, ship, and hail again. If a destination needs a roadmap to describe, it's several trips — cut it at the points where validation can happen.

6. When to stretch to six — and when to split

Three to four is the ideal. The fifth and sixth seats are headroom, not targets — legitimate while the team's AI-native maturity is still growing: onboarding a domain expert, absorbing a skill the agents can't cover yet, teaching the validation discipline itself. As maturity rises, the ideal shrinks toward the front seat.

Splitting is cheap in the agent era — the fleet scales; only the humans need re-seating. Gartner expects 60% of organizations to run smaller engineering teams by 2029 for exactly this reason (up from 15% in 2026) — restructuring around human + AI strengths, not cutting cost.

7. Anti-patterns

The one-line version

Write destinations, not code. Run agents in parallel, validate everything, watch the fare, and the moment you feel the need for a seventh seat — hail a second taxi.