Presented by:

Today's VB Daily read time: ~3 minutes.

Good morning. Google's Paradigms of Intelligence team figured out how to train AI agents to cooperate without writing coordination rules. The trick: expose agents to unpredictable opponents during RL. Instead of learning to exploit, they learn to cooperate — because cooperation is the most robust response to uncertainty. Research lead Alexander Meulemans told VentureBeat's Ben Dickson that decentralized multi-agent RL has been "notoriously difficult." PPI sidesteps the hard part. It works with standard RL like GRPO, already common in LLM post-training.

On the infrastructure side, Qdrant closed a $50 million Series B because agents need sub-100ms vector search and RAG pipelines can't deliver it. 

Below: A former Manus back-end lead open-sourced the argument that Unix CLI beats function calling, OpenAI added a shell to the Responses API, and Gemini's phone agent is ordering Starbucks.

Google finds that AI agents learn to cooperate when trained against unpredictable opponents

A message from Microsoft and NVIDIA

Know before you go: What’s new with Microsoft at NVIDIA GTC

They say it all starts here at NVIDIA GTC. Where will you start?

Explore ways to engage with Microsoft at the leading AI conference of the year. Read the Azure Tech Community Blog for all the details on Microsoft-led talks and panels, in-booth sessions, hands-on labs, demos, and more.

Make the most of every day at GTC when you prepare ahead with activities aligned to your AI goals. See what’s coming and mark your calendars for what matters most to your business.

We can’t wait to see you there.

Agents need vector search more than RAG ever did

Who's actually in charge of your AI stack?

VentureBeat is surveying enterprise technology and engineering leaders to find out what's really happening on the ground: who governs AI behavior across complex, multi-platform environments, and what's getting in the way.

Around the web

OpenAI adds a shell tool and persistent containers to the Responses API. It gives API-connected agents a full Linux environment with concurrent command execution and file persistence across turns. Context compaction keeps long sessions manageable. OpenAI is betting that agents want a terminal, not a menu of function calls.

Read more at OpenAI

Gemini's task automation beta can order your Uber and your Starbucks. Google's AI agent runs on the Galaxy S26 Ultra, navigating multi-step app workflows through natural language. It operates in a virtual window and stops before the final confirmation step — you still press the button.

Read more at The Verge

Former Manus back-end lead argues Unix CLI beats function calling for AI agents. MorroHsu says a single run(command="...") tool outperforms typed schemas. Three techniques: progressive — help discovery, error messages as navigation, consistent output via Unix pipes. Open-sourced as agent-clip.

Read more at Reddit

Random Labs releases Slate, a swarm orchestration agent for coding. Slate coordinates multiple AI models with a TypeScript DSL — routing orchestration to Sonnet/Opus/GPT-5.4 and routing execution to task-appropriate models like Codex and Haiku. Episodic memory keeps context across sessions lasting up to two days. Inspired by BEAM VM's actor model and Karpathy's LLM OS concept.

Read more on X

Moondream's segmenting update: 40% faster, 5x cheaper than SAM 3. The lightweight vision model gained +4.4 on RefCOCO+ with better mask quality and native SVG output. Worth a look if you're building vision into agents or apps.

Read more at Moondream

Percepta asks: Can LLMs be computers? Christos Tzamos explores transformers that execute arbitrary C programs using 2D attention heads, claiming exponentially faster inference for program execution compared to standard autoregressive generation. The research suggests transformers can learn to simulate computation directly rather than predicting tokens.

Read more at Percepta

Beyond the Pilot Podcast: Episode 8

LangChain told employees they cannot install OpenClaw on company laptops due to "massive security risk" — yet this unhinged approach is exactly what makes it work. Harrison Chase unpacks why OpenClaw succeeds where AutoGPT failed, and why context engineering, not just smarter models, separates demo agents from production-ready systems.

Watch Episode 8

For more on these stories and other key updates, head over to VentureBeat.com for our full coverage. 

- Bryson Masse (@Bryson_M)

For U.S.-based readers, don't miss VB on Google: Add us to your trusted feeds. Click here.

VentureBeat, Inc. · 500 Sansome St Ste 601 · San Francisco, CA 94111-3222 · USA
Privacy Policy  |  Unsubscribe from this message type