Mechanize co-founder Tamay Besiroglu has joined Google DeepMind as a research scientist, while more than a dozen former employees have moved to Google. The AI coding startup itself continues to operate under new leadership.

Google has completed a talent-focused deal with AI coding startup Mechanize, bringing co-founder and former CEO Tamay Besiroglu into Google DeepMind as a research scientist and adding more than a dozen former Mechanize employees to its AI ranks.

Many of those employees are now working on Google’s “midtraining” efforts, an area focused on strengthening particular capabilities in AI models after their initial pretraining.

The transaction has been reported at more than $1.5 billion, but there is an important qualification around that figure. The final amount and specific contract terms have not been disclosed publicly, meaning the reported price should not be treated as a confirmed final purchase value.

There is another distinction that matters. Google did not acquire Mechanize outright. The startup continues to exist, with former chief of staff Guive Assadi now listed as its CEO, while Google has secured key personnel and technology through the arrangement.

What Mechanize was building?

Mechanize’s work makes the deal more significant than a straightforward hiring exercise.

The startup has been developing what it calls a “virtual workplace” for AI agents. Instead of limiting an AI system to generating a block of code, the environment is designed to let an agent carry out more of the software development process itself.

That means writing code, executing it, identifying errors, making corrections and testing the result again. The output from that process can then be evaluated and used as feedback for training.

The approach is aimed at a familiar problem in AI coding. Producing code is only one part of development. Useful software also requires an ability to inspect existing work, understand why something failed, make a change and determine whether that change actually solved the problem.

Mechanize’s technology has been built around giving AI agents an environment in which they can go through that cycle rather than simply produce a one-shot answer.

The unusual Game Boy Advance test behind Mechanize’s approach

One of the more distinctive examples of Mechanize’s work is GBA Eval, a publicly released benchmark that asks an AI coding agent to build a Game Boy Advance emulator from scratch within a fixed period.

The benchmark is designed to measure more than whether an AI model can produce individual pieces of correct code. It tests whether an agent can take on a complicated software project and carry it through from beginning to end.

That detail helps explain why Mechanize’s work is relevant to the current push around AI agents. The challenge is increasingly not just getting a model to answer a question correctly. It is getting the system to work through a task, encounter failures, respond to those failures and keep going until the job is completed.

Tamay Besiroglu joins DeepMind as research scientist

Besiroglu, who co-founded Mechanize and previously served as its CEO, has joined Google DeepMind as a research scientist.

More than 12 former Mechanize employees have also moved to Google, with many reportedly taking roles connected to model midtraining.

Midtraining takes place after a model’s initial pretraining and is used to reinforce particular abilities. The supplied reporting identifies complex areas such as coding and mathematics as examples of capabilities that can be strengthened during this stage.

For Google, that means the Mechanize deal brings in researchers with experience in both AI coding and the environments used to train systems to handle longer, multi-step tasks.

Mechanize had raised $9.1 million earlier this year

Mechanize was founded in 2025. Earlier in 2026, the startup raised $9.1 million at a reported valuation of about $500 million.

The more than $1.5 billion figure associated with Google’s subsequent deal is therefore striking, although the comparison needs to be made carefully because Google has not disclosed the final financial terms and the transaction is not a conventional acquisition of the entire company.

Mechanize’s continuing existence under Assadi reinforces that distinction.

Google has used similar talent-focused deals before

The Mechanize arrangement also fits a pattern in Google’s AI strategy.

Google previously brought talent from AI coding startup Windsurf into the company. Windsurf’s former CEO Varun Mohan later took a leadership role in Google’s agentic coding programme, Antigravity.

The Mechanize move adds another group of researchers with specialised AI coding experience to Google at a time when major technology companies are competing not only over models, but also over the researchers and training methods needed to improve them.

Also Read: From Studying Under a Tree to a $9.3 Billion Fortune: How Jay Chaudhry Built Zscaler