DEV Community

Ashkan Rajaee
Ashkan Rajaee

Posted on

Uncountable Let 11 Engineers Go to Pay for Agent Compute. The Scarce Resource Is Still Judgment.

On November 13, 2026, Uncountable, a software company whose platform supports R&D for scientists working on new materials, chemicals and formulations, let go of 11 engineers, about a quarter of its engineering team. Unusually, it explained why in public, on a page titled "Read This First" that it keeps up for job candidates. Co-founder and CTO Jason Hirshman wrote that the company had planned engineering capacity "the traditional way: estimate product scope, derive a budget, and convert that budget into headcount," and that over the course of 2026 "that relationship broke."

The numbers on the company's engineering pages are striking. In October 2026, agents authored 91% of its pull requests, and it merged 60,449 of them at a median open-to-merge time of 19 minutes. A year earlier, with essentially the same team size, it merged 1,885 at a median of 16.5 hours. The company says its bottleneck moved from engineer hours to "available agent compute: tokens, concurrent sandboxes, and queue depth." Because model usage and salaries come out of the same operating capital, "rebalancing our capital toward agent compute meant letting good people go." It called the mistake "entirely ours" and said the decision "says nothing about their individual ability."

It is not the only signal this month. On November 9, Calcalist's CTech reported that cybersecurity company Axonius plans to cut 66 jobs, about 9% of its global workforce, as it reshapes engineering around what CEO Joe Diamond called "a new AI-driven way of working." Meanwhile, The Asia Business Daily, citing The Wall Street Journal, reported that AI startup Serval offered a $30,000 signing bonus as it hired 100 engineers this fall, and quoted Karat CEO Mohit Bende saying "the value of software engineers with strong fundamentals who know how to use AI effectively is rising significantly."

Those are the facts as reported. Everything below is my opinion.

My position: planning engineering capacity as people plus compute, not headcount alone, is overdue, and Uncountable deserves credit for explaining its decision so openly. But in a team where agents write most of the code, the scarce resource is the engineers who can judge, review and own what the agents produce. Protect and invest in that judgment before you trade people for tokens.

The budget model really is breaking

Hirshman's description of the old planning model will sound familiar to anyone who has run an engineering budget. Scope becomes a number, the number becomes heads, and output is assumed to grow with hours. I have built plans that way for most of my career. When one engineer directing agents can produce many times what they could a year ago, that math stops working.

So I agree with the core point. Compute is now a real line in the engineering budget, and it competes with salaries for the same dollars. Leaders who still treat AI tools as a small per-seat expense are likely to be surprised, either by the bill or by teams sitting in queues waiting for runs.

Where I would be careful

The risk is reading this story as "fewer engineers, more tokens" and stopping there. Uncountable's own page makes the opposite point. It says directing agents "without understanding the underlying abstractions yields fragile software," and that its durable advantage is "the technical depth and rigor of our engineering team." The company is still hiring, though far fewer people than originally planned, and with a high bar.

That is the part most teams will skip. When pull request volume jumps about thirty times, the work that does not scale is deciding what should be built, catching the subtle mistake, owning the architecture and answering for production. Agents make that judgment more valuable, not less. If you cut the people who carry it to fund compute, you get more output and less understanding of it.

Uncountable also reports that 94% of its merged pull requests had no human approval, with its own review agent checking every change. That is a deliberate choice backed by heavy investment in standards and tooling, and it may work well for them. For most companies, especially those handling customer data or money, I would keep a named human owner on every change that matters until the review system has earned that trust.

What I would do instead

Budget people and compute together. Put agent spend and engineering salaries on one plan and review the mix every quarter, so you rebalance gradually instead of in one painful cut.

Measure what matters. Track cost per merged change and defects that reach production, not pull request counts. Volume is easy to grow. Quality is what customers notice.

Hire and promote for judgment. Look for engineers who can explain why a change is right, spot what the agent missed and design the guardrails.

Keep ownership named. Every agent, service and critical workflow should have a person who answers for it.

Tell people early. Uncountable chose to explain its decision in public. Being honest about how roles are changing builds more trust than a surprise.

A fair counterpoint

To be fair, a smaller team with more compute per engineer may simply be the right shape for a company like Uncountable, and it says it is growing and still hiring. Every company's numbers and risks are different, and outsiders rarely see the full picture. The Serval report also suggests the market agrees with the core point: engineers who use AI well are in demand. My point is only about the order of decisions: decide where judgment must stay human, then decide how much compute to buy.

The payoff

Earlier in my career, as a partner, I helped grow a software consulting firm from about $6M to $60M in under three years. What scaled was never raw output. It was senior people who could make good calls and teach others to make them. Through AI Rescue & Technology Consulting, TDZ Pro provides developers, DevOps, QA, and PMs vetted over years of delivery, on a top-1% talent bench, who use AI to move faster while staying accountable for the result.

Want engineers who bring AI speed and human judgment to your team? Visit tdzpro.com.

What do you think? Should engineering budgets shift from headcount to compute, or is judgment the one thing you can never buy by the token? I would like to hear your view in the comments.

Top comments (0)