Agents Delegate. Platforms Consolidate. Intelligence Scales.

What happened
Amp announced that its agents can now spawn other agents, send messages, and exchange files across threads running locally, in remote “orbs,” or on other machines. The company framed the feature around fan-out tasks like parallel browser testing, side-quest bug fixes, and cross-project documentation work.

Why it matters
This is a meaningful product-level step toward multi-agent orchestration. Instead of a single agent working through one thread, users can now hand off sub-tasks in parallel, which is much closer to how real agent workflows will have to operate at scale.

What’s next
Watch whether this pattern spreads from coding tools into broader work agents. Once delegation becomes visible and reliable in the UI, orchestration stops being an architecture diagram and starts becoming a user-facing feature.

What happened
At WAIC in Shanghai, Xi Jinping cast China as the champion of a new global AI order, promoted open-source AI as a public good, and pledged AI training and cooperation for countries across BRICS, ASEAN, Latin America, and Africa. Reuters reported that Xi tied that message to the launch of the China-backed World AI Cooperation Organization and positioned Beijing as an alternative to Washington on AI governance.

Why it matters
China is no longer treating open models as just a developer adoption tactic. It is using them as geopolitical leverage, which means the AI race is increasingly about standards, alliances, and distribution—not just benchmarks.

What’s next
Watch the first U.S.-China government-level AI talks and whether more countries align with Beijing’s framework. If they do, open-weight AI becomes part of statecraft, not just product strategy.

What happened
Databricks disclosed a new funding round at a $188 billion valuation, as the company continues its shift from big-data backbone to full AI platform with products including Lakebase for AI agents, Unity AI Gateway, and Omnigent for managing multiple agents.

Why it matters
The market is rewarding companies that own the enterprise data layer, governance layer, and agent tooling layer all at once, which says a lot about where value is likely to accrue in the next phase of AI adoption.

What’s next
Expect the platform fight to intensify as vendors compete to become the system enterprises trust to run agents on top of their data, not just the place they buy model access.

What happened
Nvidia positioned in a blog post that post-training, not pretraining, is becoming the central workload of the agentic era, and said Vera Rubin is designed to maximize “intelligence per dollar” by making continuous reinforcement-learning loops cheaper and more scalable; the company also pointed to Nemotron 3 Ultra scoring 71.7% on SWE-bench Verified.

Why it matters
That framing matters because it shifts the infrastructure conversation away from raw chip counts and toward whether a stack can continuously improve models after deployment, which is exactly the problem agentic systems create.

What’s next
Watch for hyperscalers, neoclouds, and enterprise AI platforms to copy this language fast, because post-training efficiency is becoming the cleaner way to justify AI spend once benchmark inflation stops impressing buyers.

What happened
Agility Robotics is opening a 60,000-square-foot facility in Fremont, California, near Tesla’s expected Optimus manufacturing base, and says its Digit robot is already generating revenue in manufacturing and warehouse settings for customers including Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada; the company says it has secured $300 million in contract orders.

Why it matters
This is one of the clearer signs that physical AI is moving from prototype theater to operational rollout, with a humanoid robotics company talking less about future possibility and more about live deployments, safety integration, and contracted demand.

What’s next
The next milestone is volume and reliability: if Agility can turn pilot-style deployments into repeatable production wins, Fremont could become a real proving ground for commercial humanoids rather than just another symbolic robotics address.

What happened
WeRide unveiled WITT, a physical AI cognitive foundation model that converts real-world driving data into what it calls “Atomic Physical Facts” for extraction, reasoning, verification, and curation. The company says WITT is designed to turn operational data into trusted signals for training and evaluation, and claims up to 98% lower token costs and up to 200x greater data-processing efficiency than much larger general-purpose models.

Why it matters
This is a notable bet that physical AI will be won less by a single giant embodied model than by a better pipeline for turning messy real-world events into usable learning signal. It also reinforces the idea that autonomous driving may become the first scaled proving ground for physical-AI infrastructure before humanoids do.

What’s next
The big question is whether this fact-based approach transfers beyond driving into broader robotics and industrial workflows. If it does, the bottleneck in physical AI may shift from hardware spectacle to reality-grounded data operations.

The next phase of AI is about systems, not models. The winners will orchestrate agents, improve continuously, and control the platforms where intelligence learns, works, and scales.

Try building a manager for your AI agents.

Use Amp, Claude Code, or OpenAI Codex to break a project into parallel tasks. Have one agent research, another write, another review, and then merge the results.

The biggest productivity gain often comes from coordinating multiple agents instead of asking one agent to do everything.