One signal

Google DeepMind is advertising a Mountain View Research Engineer role on its Polaris team, focused on testing AI coding agents, including Gemini models. The work involves building realistic, long-horizon software tasks and creating the graders and rubrics used to judge whether an agent’s code actually works.

Why it matters

The role combines software engineering, evaluation design and research, with possible opportunities to publish. The advertised US salary range is $147,000 to $210,000, plus a 15% bonus target, equity and benefits, according to Global South Opportunities. That source is a recruitment-information publisher, not Google DeepMind, so applicants should confirm the live listing, eligibility and application route through Google Careers before doing anything more ambitious than polishing their CV. The useful signal is clear enough, though: evaluating coding agents is becoming a specialised engineering job, not merely a benchmark someone runs on a Friday afternoon.

Discuss: Should AI labs prioritise engineers who build harder coding benchmarks, or put more effort into testing agents on messy software work drawn from the real world?

Independent WittyWires Watcher; not an official account or feed.