AI
Agents Talk to Each Other as Open Weights Keep Flowing
The most important AI story this week is easy to miss if you watch only for a smarter chatbot. It is the plumbing. Google marked agent to agent communication stable, Anthropic taught its coding agent to run tool calls in the background, and a flood of open weight models arrived from Tencent, Moonshot, and AMD, all in the same few days. The pattern is plain. Intelligence is moving out of the chat window and into systems that call each other, check each other, and keep working while humans do something else.
Google shipped ADK 2.14.0 on November 7, moving its agent to agent protocol, known as A2A, plus execution environments and the core runtime, to stable. Stable is the word that matters. Teams can now build agents that hand work to other agents with a protocol they can count on staying put. The same day, Anthropic shipped Claude Code 2.1.319, which runs MCP tool calls in the background instead of making the user wait on each one, adds a setting called denyInsteadOfPrompting, and changes how its loop command behaves. The two releases rhyme. One standardizes how agents talk to each other, the other stops the coding agent from pausing the human every time a tool call takes a while.
The model flood kept pace with the plumbing. Tencent shipped Hy4 on November 4, a 770 billion parameter mixture of experts model with native visual reasoning and computer use, a 1 million token context window, and open weights, priced near 90 cents per million input tokens. Moonshot AI followed on November 5 with Kimi K3.1 Pro, a 2.8 trillion parameter mixture of experts model with 104 billion active parameters, open weights under a modified MIT license, and a top ranking among open models on Agent Arena. AMD joined the same wave from an unexpected direction, with Instella 3B SFT, a small model trained from scratch on AMD Instinct MI250X graphics cards and released with weights, code, and data recipe under Apache 2.0, a proof that competitive training runs on hardware outside the Nvidia stack. November now counts 20 model launches from 18 vendors, with half offering open weights.
Security got its own headline the same week. OpenAI introduced GPT 6 Cyber on November 5, a cybersecurity model for authorized vulnerability discovery, available in limited access and carrying the first Critical rating under the company's Preparedness Framework. It scored 48.5 percent on ExploitGym and completed 95 percent of tasks on a benchmark for remote attacks, priced at 20 dollars per million input tokens and 100 dollars per million output tokens. OpenAI also pointed Hacktron, a security firm, at its own systems and integrations to hunt for vulnerabilities, a move that treats outside eyes as a feature of the build. On the creative side, OpenAI split its image model in two with Images 2.6. Flare runs fast for drafts, Sunburst takes the careful path for quality and consistency across repeated edits, generation runs about 50 percent faster, and the release adds GIF creation and SVG export.
One more release deserves a careful read. Thinking Machines Lab shipped Inkling 2 Small on November 5, a 594 billion parameter mixture of experts model at 1.18 dollars per task, and many outlets called it open weights on arrival. The weights are still to come. The company promised an Apache 2.0 release after outside safety testing, with the release date still unannounced, and the model runs only on its Tinker platform today. The distinction matters because the open weight label shapes how builders plan, and a promise is different from a download. The week still counts as a strong showing for open weights, and the gap between promise and download is the detail to watch.
For readers, the takeaway is practical. Agents that talk to each other and keep working in the background turn AI from a tool you prompt into a colleague you brief. Open weights from Tencent, Moonshot, and AMD put that colleague within reach of smaller teams, and the security work from OpenAI shows the frontier labs treating their own systems as something to test in public. The next thing to watch is what builders ship on stable A2A in the coming weeks, because a standard only proves itself when real work flows across it.
Quick answers
What is this story about?
The most important AI story this week is easy to miss if you watch only for a smarter chatbot. It is the plumbing. Google marked agent to agent communication stable, Anthropic taught its coding agent to run tool calls in the background, and a flood of open weight models arrived from Tencent, Moonshot, and AMD, all in the same few days. The pattern is plain. Intelligence is moving out of the chat window and into systems that call each other, check each other, and keep working while humans do something else.
Why does this story matter?
For readers, the takeaway is practical. Agents that talk to each other and keep working in the background turn AI from a tool you prompt into a colleague you brief. Open weights from Tencent, Moonshot, and AMD put that colleague within reach of smaller teams, and the security work from OpenAI shows the frontier labs treating their own systems as something to test in public. The next thing to watch is what builders ship on stable A2A in the coming weeks, because a standard only proves itself when real work flows across it.
Sources
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.