Weekly AI Roundup (November 16, 2026): Gemini Decision 1, self-improving agents, and the open-weight wave

In one week, AI moved from “models that write” to models that decide, and from “agents that execute” to agents that improve themselves. The essentials, for decision-makers.

Decisions, not text: Gemini Decision 1

Google launched the Gemini Decisions API with its gemini-decision-1 model, built for deterministic control planes, not chat. Instead of generating text token by token, it scores up to 1,000 options in a single pass and returns probabilities with a 0-to-1 confidence score.

For your company: agents stop “thinking out loud.” Classifying a ticket, approving or rejecting a case, choosing the next action — all with a bounded, auditable output. Exactly the kind of piece an agent needs to run in production without hallucinating.

Open weights: ERNIE 5.5, Praxis-1, and Bielik

• Baidu ERNIE 5.5 Flash (280B, 1M context) wrote a full Unreal Engine 5.8 game in a single 11-hour run. Apache 2.0 weights promised by November 25.
• Runway open-sourced Praxis-1, a 27B model that predicts video and robot actions together, topping NVIDIA’s RoboLab-120 leaderboard.
• Bielik 11B v3.1 added function calling, structured output, and decision modes — an open-weight built for agents.

The pattern is clear: yesterday’s closed frontier becomes downloadable. For sensitive-data use cases, self-hosting stops being a luxury.

Enterprise: agents that improve themselves and gain identity

• Salesforce published SFR-AutoR&D: agents that run the full research loop and improve the models the company already runs in production (+36% throughput, −17% latency on embeddings).
• AWS shipped AWS Emulator with a striking detail: 200+ services emulated locally, written entirely by thousands of AI agents.
• Linux Foundation launched Agent Name Service (ANS): verifiable agent identities via DNS and certificates — a provenance control you’ll soon be able to demand in procurement.

Money: two big bets on the agent era

• Fractile raised $720M at $7B for memory-first inference chips optimized for the long contexts agents create.
• Rox became a unicorn ($1.3B) with “revenue agents” that research accounts and prepare outreach.

What to take away

1. “Deciding” is the new “writing”: add decision models for classification, triage, and approvals with explicit confidence.
2. Open weights now compete seriously: re-evaluate self-hosting for sensitive data before renewing licenses.
3. Agents need identity and audit: demand provenance and traceability from any vendor before connecting them to your systems.

New roundup every Monday. RSS at /en/rss.xml. Applying AI to your operation? Let’s talk about your case.