Interactive Lab · Agentic engineering · Amdahl's law

Coding just got 10× faster. Why does the feature ship only a little sooner?

Writing code is one stop on a longer route: someone has to decide what to build, design it, review it, test it, ship it, and a lot of the time the work just waits in a queue. AI coding agents speed up one stop. Amdahl's law says how much that helps the whole trip: the parts you didn't speed up set a hard ceiling. Drag the sliders and see it for yourself.

▶ Predict first, then drag

🔮 Commit a guess first

In this example, coding is 25% of the route. If coding gets 10× faster, how much faster is the whole route?

Pick one. Then drag coding to 10× and check your guess.

Whole route1.00×faster end to end

Lead time10.0 dwas 10.0 working days

Coding-only ceiling1.33×even if coding took 0 time

Coding speed-up 1×

1× · as todayAI agents · 20×

Coding's share of the route 25%

5%60%

Speed up the whole route

Now speed up the other stops too: faster reviews, automated tests, less waiting between hand-offs.

The stage shares are an illustrative example of one team's feature, not measured data. Use the share slider to match your own team.

🧭 Where this fits

Gene Amdahl wrote this law in 1967 about computer processors: speed up part of a program and the rest limits you. The same arithmetic applies to a team. When AI agents make writing code cheap, the slow stops become review, testing, decisions and waiting. Teams that only speed up coding get a small win. Teams that speed up the whole route (smaller changes, automated checks, fast reviews, fewer hand-offs) get a big one. This is the delivery side of DevOps and CI/CD.

The code that's actually running

🔮 Predict, then test

Set coding to 10× and leave the other stops at 1×. Then try each of these alone: coding at 20×, waiting at 2×, deploy at 5×, design at 2×. Which one moves the whole route most? The answer is always the stop that now holds the most time.

The takeaway

Your new mental model

Amdahl's law: speed-up = 1 ÷ (the part you didn't speed up + the part you did ÷ how much faster). The slow part sets the limit.
If coding is 25% of the route, even infinitely fast coding gives at most 1.33×. Making coding 10× faster gives about 1.29×.
Once coding is fast, the bottleneck moves to review, testing and waiting. That's where the next hour of effort pays off.
To ship much faster, speed up the whole route: automated checks, small changes that are quick to review, and fewer queues between people.

Next: the agent loop and the review queue →Try: CI/CD pipelines All playgrounds