AI digest: Agents go mainstream
Big tech companies are finally shipping agent platforms while new open source models challenge the frontier labs.
This week was all about making AI agents actually useful for normal people, not just demos.
OpenAI turns ChatGPT into a proper automation platform
OpenAI launched workspace agents that run continuously and automate team workflows, powered by Codex. This is the first time we’ve seen a major lab ship agents that actually work in the background without human babysitting. Custom GPTs are sticking around for now, but this feels like the real future of workplace AI.
Google goes all-in on enterprise agents
Google unveiled its Gemini Enterprise Agent Platform alongside 8th-gen TPUs at Cloud Next. Interesting choice to target IT teams rather than business users directly. They’re also adding AI overviews to Gmail and letting Chrome auto-browse for enterprise users. Feels like they’re trying to own the enterprise agent stack end-to-end.
Xiaomi’s MiMo models match frontier performance for less
Xiaomi released MiMo-V2.5-Pro and MiMo-V2.5, claiming they hit frontier model benchmarks at much lower token costs. If true, this is huge for making capable agents actually affordable to run. The open source ecosystem is moving fast enough that the closed model advantage might be shrinking faster than expected.
Qwen3.6-27B outperforms much larger models
Alibaba’s new Qwen3.6-27B model is apparently beating 397B mixture-of-experts models on coding benchmarks whilst being much smaller. Dense models making a comeback suggests the scaling laws might be more nuanced than just “bigger equals better”. Good news for anyone who wants to run capable models without massive compute budgets.