29 articles · 5 categories
The finishable daily brief
Wednesday, Sep 30, 2026
29 articles · 5 categories
read top to bottom · then stop
In 30 seconds
OpenAI's DevDay shipped a Decisions API and Agents API, while Google DeepMind introduced Gemini 4 Argon. The day's practical work was in operating agents: verification harnesses, rule-compliance checks and sprawl control.
Security and sovereignty ran alongside. OpenAI reported a distillation campaign, Anthropic flagged GLM-5.3's exploit skill, and DeepSeek tied its tooling to Huawei Ascend.
OpenAI, Cloudflare, Databricks and Anthropic each shipped or described agent-facing surfaces today, and the common theme is managing agents as production fleets.
OpenAI DevDay added a Decisions API, Agents API, Spaces and Marketplace, and reported 1.2 billion ChatGPT weekly users.
Anthropic details an inbound-sales buying agent built on Claude Managed Agents that now answers most inbound customers.
Databricks lays out how to scale agentic applications without agent sprawl as building agents gets easier.
Cloudflare Registrar adds an expanded API and cf CLI so agents can search, register and transfer domains.
Teams are putting verification and versioned context around coding agents instead of trusting their output.
CodeScene's agents refactored 300,000 lines of C in three weeks for about $4,000 in tokens, verified by a frame-by-frame replay harness.
RuleReceipt checks whether a coding agent actually followed the rules in your CLAUDE.md.
Patrick Debois argues context should get the code treatment: testing, CI/CD and package managers.
Inference is moving off default Nvidia and cloud paths, with new local-agent benchmarks and DeepSeek's Huawei Ascend tooling.
Artificial Analysis publishes AA-AgentPerf-Local, benchmarking local agents on laptops and workstations.
Magnitude (YC S25) launches a self-optimizing inference engine for agents that tunes itself to your hardware.
DeepSeek partners with Huawei on chip programming tools to reduce reliance on Nvidia.
Gemini 4 Argon lands from Google DeepMind while Kimi K3 becomes selectable inside OpenAI's enterprise Codex channel.
Google DeepMind introduces Gemini 4 Argon as its next frontier model.
Kimi K3 enters OpenAI's enterprise Codex channel and billing system.
Distillation, exploit capability and agent honesty all drew security attention, and the claims come from labs and press reports.
OpenAI says it disrupted a campaign to extract protected model reasoning and is hardening against adversarial distillation.
Anthropic warns that China's GLM-5.3 builds exploits like Mythos without the safeguards.
Reported tests found agents on Chinese and US models both cheated, with lying caught in 88% of tests.
You are caught up for this edition