Directed velocity is the constraint.

AI changes what is scarce. The ability to turn a specification into working software is rapidly becoming abundant. The ability to choose the right specification—and to continuously test it against reality—is not.

That shift changes the job of an engineering organization. Output is no longer a sufficient proxy for progress. A team can produce more code, more agents, and more surface area while quietly moving further from the customer problem.

Velocity is becoming a commodity

For most of software history, implementation capacity was the dominant constraint. We designed organizations around allocating it: roadmaps, tickets, estimates, and layers of coordination. Agentic development compresses that loop. The constraint moves upstream to intent and downstream to validation.

Speed without a shared vector is simply a faster way to distribute debt.

“Directed velocity” is the useful output remaining after alignment errors, coordination drag, and systemic debt are accounted for. It is not a call to slow down. It is a demand that speed remain coupled to a goal.

A useful model

Consider each team or agent as contributing a velocity vector $v$, scaled by the quality of its context $q$. Project the combined effort onto the goal direction $g$, then subtract the debt accumulated by the system:

The exact numbers are less important than the discipline of the model. It forces us to ask whether faster local motion improves global progress.

Architecture for directed systems

A healthy multi-agent system needs more than specialized workers. It needs a shared state of truth, explicit evaluators, and an observer capable of updating the plan when reality disagrees.

The operating ratio

The simplest metric I return to is a ratio: how much validated learning did the system create for each unit of complexity and coordination?

1. 01
   Make the goal executable.

   Translate ambition into tests an agent, engineer, and customer can agree on.

2. 02
   Route context, not just tasks.

   The quality and timing of context determine the usefulness of generated output.

3. 03
   Instrument for learning.

   Every release should reduce uncertainty, not merely increase surface area.

4. 04
   Price the debt immediately.

   AI can produce complexity faster than organizations can understand it. Keep the cost visible.

The competitive advantage is not AI-speed by itself. It is the ability to aim that speed, observe its effects, and change direction without losing coherence.