The Signal · Column 105

The end of tokenmaxxing makes engineering judgment the bottleneck

Andrei van Noordt · 5 November 2026 · 2 min read

Sifted's outlook on engineering in 2026 compresses the year into a single line: fewer people, more work, and the end of tokenmaxxing. It is a forecast, not an announcement, but it puts words to something I hear in almost every hiring conversation this autumn. Teams are being asked to ship more with smaller headcount, and the novelty of simply consuming more AI is wearing off.

The part most people will skim past is the last phrase. Tokenmaxxing, as I read it, is the habit of treating AI usage itself as progress: more prompts, more agents, more generated code, more spend on the meter. Its end does not mean companies are cooling on AI. It means they have started asking what the output is actually worth. Once that question is asked, the bottleneck stops being the model and becomes the person deciding what to build, what to keep and what to throw away.

When volume stops being the metric, judgment becomes the scarcest resource on the team.

That is the repricing. A smaller team doing more work cannot carry passengers, and it cannot carry people whose main contribution was typing code the tools now produce in seconds. What it needs is engineers who can scope a problem, read generated output critically, and own the consequences in production. Those profiles were always rare. In a team of eight, every one of them is load bearing, and the gap between a strong hire and an average one widens with every tool you add.

It also changes the junior question. If fewer people do more work, the entry level does not disappear, but it becomes harder to justify unless someone senior has the time to turn a junior into the kind of engineer the tools amplify rather than replace. Companies that skip that investment will find themselves competing for the same small pool of experienced people later, at a higher price.

For hiring leaders, the practical read is simple. Stop measuring your engineering organisation by headcount or by how much AI it consumes, and start measuring it by how much sound judgment sits on each team. Hire for that first, and let the tools take care of the volume.