When AI Writes the Code, What Becomes the Bottleneck?
The people who can judge what comes back. When AI writes more of the code, the speed of a team is set by how fast someone can review it and decide whether it belongs in the product. Across more than 500,000 developers, autonomous coding agents raised commits by 240%, and releases by only 30% (NBER, 2026).
Engineers at Anthropic report that their work is also moving towards directing agents and deciding what to build; that shift has not been measured in hours yet. Tunga staffs experienced digital professionals from across Africa, and its selection tests whether they work with AI tools and can review code an agent wrote. You can add your own test on your own code before you choose.
What the measurements show
- Much more code, much less extra shipped work. NBER Working Paper 35275 (Demirer, Musolff and Yang, 2026) follows more than 500,000 developers on GitHub through three generations of AI tools. With autonomous agents, commits rose 240% and releases 30%. In four app stores the number of new apps rose sharply while total use stayed flat. The authors' explanation: writing got faster, but the steps around it still run at human pace, so the gains "have therefore translated only partially into shipped and used software thus far."
- Pull requests wait longer for review. Faros AI's AI Engineering Report 2026 covers more than 4,000 teams (22,000 developers). At each organisation's highest AI use, the median time a pull request spent in review was 441% higher than at its lowest. Faros sells engineering analytics, so this is vendor data; the independent NBER study points the same way.
- When review can't keep up, it gets skipped. In the same Faros data, pull requests merged without any review rose 31% and bugs per developer 54%. That is a correlation within organisations, not proof that AI causes bugs.
- More of the code starts with an agent. By September 2026 more than 40% of pull requests on GitHub started from an agent's first draft (GitHub Octoverse 2026).
What engineers say moves next
One engineer at Anthropic described their work as moving "70%+ to being a code reviewer/reviser". In Anthropic's study of its own engineers (132 surveyed and 53 interviewed, 2025), most could fully hand off no more than a fifth of their work. They hand off what is easy to check, and keep design and the decisions that need taste or context. Others describe themselves as managers of AI agents. That shift to directing and deciding is self-reported, at an AI lab, and nobody has measured it in hours yet. What the same shift means for work beyond software is in the blog If you mostly work at a computer, this is how your job will change.
What this means for your team
- Who reviews matters more than who writes. Judging agent output takes someone who knows the codebase and can turn down a pull request that works but doesn't fit.
- Part of the check is moving to AI as well. Use of AI-assisted code review on GitHub grew more than 90% in a year (GitHub Octoverse 2026). The decisions that are expensive to get wrong stay with a person.
How Tunga adds people who can judge what comes back
- The skill this page is about is tested before anyone reaches your shortlist. Every professional has at least 3 years of experience and passes technical tests and a review of their own code on GitHub. The selection also tests whether they work with AI tools and can review code an agent wrote (see how Tunga selects professionals).
- You can add your own test on top. Some clients add their own test, and Tunga sets it up with the shortlisted candidates, so you see that judgment on your own code before you choose.
- You hear early whether it works. The delivery manager calls the professional at the end of week 1 and checks in with both sides at least once a month, so a gap between the judgment you need and what the person brings comes up in one of those conversations.
Where Tunga fits
Tunga takes the staffing side off the plate of companies that need digital professionals: software engineers, AI operators, product managers and other digital roles, from across Africa. The client directs the work. Tunga handles finding and vetting, contracts, payments and compliance, onboarding, keeping people, and replacing them when someone leaves or it isn't working. The professionals are independent contractors or employed in their own country through an employer of record, and work remotely from their own country; the client has one contract, with Tunga, and a delivery manager who checks in with both sides. In 2025, about 1 in 56 placements ended because the professional left for another job. Founded in 2015; 3,500 professionals across 38 African countries and 300+ client companies, as of September 2026.
Tunga is not the right fit when the role has to be on site, when the work means talking to your clients in a language other than English or French, or when saving on cost is the main reason you're looking.
Need capacity now?
Tell us what you're looking for. We'll get back to you within one working day.
Tell us what you need