Agents Take Control. Coordination Becomes the Product.

What happened
OpenAI's Codex can now control Windows computers, meaning the app can see the screen and take actions on-device. OpenAI is also allowing users to review Codex jobs directly from the ChatGPT app.

Why it matters
This pushes coding agents beyond terminal windows and into full desktop workflows, which is where a large portion of enterprise work still happens. It also raises the bar for agent platforms from "can write code" to "can operate the environment."

What's next
Expect the coding-agent race to shift toward orchestration, review, and computer-use workflows rather than raw autocomplete. As agents gain access to more tools and environments, execution may become more important than generation.

What happened
Asana acquired Stack AI, a no‑code agent builder, for $75 million—its first acquisition in 18 years—alongside an earnings beat that briefly lifted its stock. The move comes after Asana lost roughly half its market value during the AI boom and aims to reposition the work‑management company as the coordination layer between human workers and AI agents.

Why it matters
Asana’s seat‑based SaaS model is under pressure as AI agents can do the work of multiple humans, prompting a $1 trillion wipe‑out in SaaS market capitalization. By focusing on agent orchestration, Asana hopes to capture new revenue streams—its AI products already generate over 17% of new ARR and customers spending more than $100K on its AI tools have nearly doubled.

What’s next
The company plans to integrate Stack AI within months, giving it cross‑system automation capabilities that rival Salesforce and ServiceNow. CEO Dan Rogers predicts most employees will soon work alongside agents, so Asana’s success will depend on solving the coordination challenge and persuading investors that its business model can still scale.

What happened
Cognition CEO Scott Wu told TechCrunch that Devin now commits 89% of the code written by Cognition engineers, with the agent handling much of the long-tail maintenance work inside the company.

Why it matters
This is one of the clearest public signals yet that coding agents are moving from showcase tasks to real internal throughput. The conversation is shifting from "can agents code?" to "how much software can a team safely hand over?"

What's next
Vendors will face increasing pressure to show production evidence—not just benchmark wins or fundraising momentum. Customers will want proof that agents can reliably contribute to shipping software.

What happened
Business Insider reported that companies are blowing through token budgets as agents drive up usage. In response, Google is positioning Gemini 3.5 Flash as a lower-cost alternative to premium frontier models. Google says large cloud customers could save more than $1 billion annually by shifting workloads to Flash and related offerings.

Why it matters
The competitive axis is moving from pure model quality to inference economics. As agents become more capable—and more expensive—cost control is becoming a product feature rather than just a finance problem.

What's next
Expect more enterprises to route workloads across multiple models instead of standardizing on a single provider. That favors vendors with lower infrastructure costs, better routing systems, and stronger cost visibility.

What happened
Silicon Angle reported that XCENA raised $135 million at a $570 million valuation to build chips that move routine AI data processing closer to DRAM. The goal is to reduce costly movement between CPUs, GPUs, and memory systems.

Why it matters
This is a bet that the next AI infrastructure battle will be won below the model layer. If inference is increasingly memory-bound, the highest-value optimizations may come from system architecture rather than marginal model improvements.

What's next
Expect more capital to flow into the plumbing around inference—memory, caching, interconnects, and orchestration—as hyperscalers search for savings at scale.

What happened
Microsoft moved Copilot Health into preview for U.S. Microsoft 365 Personal, Family, and Premium subscribers age 18 and up. The product allows users to combine health questions, wearable data, and health records into a single Copilot experience following what Microsoft described as extensive safety testing and evaluation.

Why it matters
AI products are moving deeper into regulated and sensitive workflows where trust and governance matter as much as model quality. Healthcare represents a significantly harder proving ground than productivity software, making this an important test of how far consumer AI assistants can extend.

What's next
If the preview performs well, expect more AI products to evolve from general assistants into domain-specific copilots with stricter safety, compliance, and reliability requirements.

What happened
The Verge reported that startups are increasingly paying people to perform household chores while cameras and sensors record every movement. High quality real world data has become a major bottleneck for training robots and other physical AI systems. One example is Shift, which offers free home cleaning in exchange for training footage.

Why it matters
Physical AI is running into the same constraint generative AI once faced with web-scale data. The challenge is no longer just building better models—it's obtaining the right training data. The robotics data moat may ultimately belong to whoever can capture messy human motion at scale rather than whoever produces the most impressive demo.

What's next
Expect more "data-for-service" business models in robotics, along with increasing scrutiny around privacy, consent, and labor practices.

What happened
BC Hydro is deploying four‑legged robots—Spot, Spark and Bolt—to inspect substations, patrol facilities and access hazardous areas. The utility is also developing a bipedal humanoid robot for indoor and underground work, and plans to integrate AI for automated acoustic and visual inspections.

Why it matters
The robots reduce human exposure to dangerous environments while improving the reliability of British Columbia’s electricity system. Spot and Spark can climb stairs, open doors, read gauges and detect hotspots, while Bolt’s agility allows it to navigate tight underground spaces, making BC Hydro’s program one of the most advanced in the utility sector.

What’s Next
After expanding its robotic fleet in 2025, BC Hydro will test the humanoid robot indoors and incorporate AI-powered inspection capabilities. If successful, the approach could become a model for utilities worldwide seeking to enhance safety and efficiency.

The AI race is moving from intelligence to execution. Agents are gaining access to desktops, workflows, health records, software teams, physical environments, and enterprise systems. As capabilities improve, the biggest differentiators will be orchestration, cost efficiency, trust, and operational control—not just model quality. The organizations that learn how to supervise, coordinate, and govern fleets of agents will have a larger advantage than those simply deploying the smartest models.

Build a simple multi-agent workflow and see where the bottlenecks actually appear. Use OpenAI Codex or Claude Code to tackle a real project—such as building a landing page, automating a report, or analyzing a dataset. Then break the work into specialized roles: one agent for research, one for coding, one for testing, and one for review.

Next, experiment with orchestration. Use a no-code platform like Stack AI or a workflow tool such as Asana AI Studio to coordinate tasks across multiple agents and compare the results to a single-agent workflow.

Finally, track the economics. Run the same workload on a premium model and a lower-cost model such as Google Gemini Flash and measure the tradeoffs in speed, quality, and cost. As today's stories suggest, the future advantage may come less from having the smartest model and more from knowing how to route, supervise, and coordinate fleets of agents efficiently.