Widen the highway.
Now watch the exit.
Start by boosting coding alone — that is the AI-first pitch. Watch where the cars pile up. Then add product and QA capacity, and finally give Review room. Each car is a task; the result that matters is how many reach customers. All simulator numbers are model assumptions, separate from the research below.
Tasks reaching customers, across the whole route.
Felt speed: 3× coding capacity. The highway has room for 24 tasks/week; only 10 reach it and get coded.
- 01
Old town
Discovery, analysis, prototype
Spec & prototype10 / week capacity120 waiting before this stage - 02
The highway
AI-assisted coding
Code24 / week capacity0 waiting before this stage - 03
The toll booth
Read, challenge, security check, approve
Review12 / week capacity0 waiting before this stage - 04
The exit
Test the whole journey
QA / verify8 / week capacity24 waiting before this stage
Moving cars = work processed. Parked cars = waiting tasks (up to 18 drawn per queue). Lane counts and driving speed are illustrative; the capacity numbers drive the calculation.
96 delivered so far · 120 ideas waiting to enter · 24 tasks waiting inside the town (WIP)
QA / verify sets the pace. Turn up coding alone. Then widen the streets around it.
How this model works
One task, one route
Every car is an equal-size task. Empty queues at the start; 12 weekly steps; same-week handoffs. No rework, batching, context switching or delivery delay is modeled. This is flow arithmetic, not a road-traffic or delivery forecast.
The arithmetic
Code capacity = 8 × AI boost. At each stage: handled = min(queue + arrivals, capacity). Unhandled work stays queued. Delivered/week = min(ideas in, Spec, Code, Review, QA).
Count the whole queue
Ideas awaiting Spec are the incoming backlog. Work that has passed Spec but is waiting for Code, Review or QA is WIP. Total waiting = both. Adding capacity may grow WIP at the next constraint even as total waiting falls. It cannot guarantee an empty queue.
AI-first is the road.
The town is the product.
Three streets that fast coding does not widen. On the left, what a coding-only speedup looks like from the inside. On the right, what the same work needs once it has to reach a customer — the part that is easy to underestimate when the code is arriving so quickly.
Know what to build.
Read it. Then check it.
Prove it works for a user.
We built the town around our own highway.
Our engineers have been AI-first for two years. What changed in 2026 is who we hire: mostly testers and product people. Not because coding got worse — because everything around it became the constraint. Four things we now say to anyone who tells us their developers got faster.
Discovery and analysis are where speed is decided.
A highway only helps if the cars are going somewhere worth going. Product people turn requests into shared understanding before code starts, and they cut the features nobody needed — the cheapest speedup there is.
Review and security checks are a real cost. Budget them.
Generated code is long, fluent and confident. Reading it with comprehension and checking it for security — permissions, injection, secrets, failure paths — is slow by design. Give that work a visible queue, an owner and time. Pretending it is free is how incidents happen.
Testers are the exit lane.
Users move between screens, roles and devices. Tests have to cover those journeys, including failed requests and unexpected input. Our QA people plan and run that verification so faster development has somewhere to go — it is the role we hired for most this year.
A plausible answer can still miss the need.
On Elite Medical Prep and a real estate product the model suggested things users did not want and asked questions too broad to help. Nobody fixed that with a bigger model. Product people defined useful behavior, testers checked it against real scenarios.
More activity is one signal.
Delivery is another.
These sources measure different things, in different settings. They motivate the questions in this brief; they do not calibrate the traffic model or predict your team's results. The hero figures come from the linked Faros study and from our own hiring.