It's striking how far we've come since the early days of AI "function calling" in June 2023. While embedding API calls within LLMs—and later MCP—unlocked a wave of agentic capabilities, it also quickly exposed structural limitations: repeated round-trips, token overhead, fragile JSON schemas, and ever-growing orchestration complexity.
Interestingly, we now see two prominent actors adopting an intuition that Hugging Face was the first to formulate more than a year ago with their "smolagents": let the model write code, run that code in a sandbox, move orchestration out of the context window.
Recently, both Anthropic (see "programmatic tool calling" paper, Nov. 2025) and Cloudflare ("Code Mode", Sept. 2025) seem to be pushing toward more execution at the edge, which opens potentially divergent strategic options among the major industry players. One can re-read each of their latest announcements as a battle not merely to control the model, but to command the very runtime that executes it.
For my part, I see several possibilities emerging.
1. Edge-First Runtimes (e.g. Cloudflare's Workers)
After securing, accelerating and distributing the web—remember their "pay-per-crawl" idea and the newly announced "markdown for agents"—Cloudflare is taking a step toward an era of distributed agents, promoting a model-agnostic runtime at the edge.
Agents as network primitives.
The idea clearly stands: browser structured the web 1.0, cloud structured the web 2.0, mobile opened the web 3.0. Could agentic runtime structure, tomorrow, a mostly AI-purposed web?
2. Cloud-Hosted Sandbox Containers (AWS)
For cloud-first historic players like Amazon Web Services, agents should live just where other apps live: in the enterprise cloud. To enable it, they leverage a particularly comprehensive, model-agnostic toolkit that spans everything from compute to containerization, orchestration, and observability—all integrated seamlessly.
Agents as cloud workloads.
3. Browser-Based Intelligence (Chrome)
AI-assisted browsers have become many these past two years—Comet, Dia, Atlas, Chrome with Gemini—but the WebMCP spec shipped in beta with Chrome 146 could be an even stronger move: not only equipping the browser with native AI capabilities, but trying to trigger a broader standard shift at the publisher level (remember when Google also tried with the AMP news format). In this vision, the browser could become "the agent runtime"—as well as a main access point to user data—a situation where Google would monetize its unrivaled distribution surface.
Agents as a new layer of the browser.
At this stage, it is worth noticing that Google operates simultaneously across these three integrated layers: cloud runtime (Vertex and its enterprise agents), browser/OS (Chrome/Android), as well as keeping Gemini a multi-purpose model with built-in execution—native tool and function use, multimodal reasoning and action orchestration.
4. Locally-Distributed, On-Device Intelligence (Apple)
Staying true to its long-standing position, Apple remains dominant in the "personal-edge" arena, enabling new productivity apps—such as Claude Code and the Codex App—to achieve spectacular adoption breakthroughs with embedded intelligence, whether powered by cloud services or run entirely on-device. Local runtime brings sovereignty and confidentiality, and opens an interesting alternative for enterprises seeking strategic independence—a counterbalance to the often-captive all-cloud ecosystems. This looks like distributed intelligence at the edge of the enterprise.
Agents as an extension of the device.
5. Agents in the Application Layer (Microsoft)
Halfway through a precarious mutation of its traditional business model, Microsoft doesn't take a clear position in the shift of the AI runtime's center of gravity—veiling it behind the façade of an app-centric, "agentic enterprise platform." Their massive enterprise userbase, combined with the unique OpenAI deal, may have brought Azure very strong traction until now. Yet that alone does not offset the hazy execution that persists across the M365 application layer, whether for personal and enterprise use.
6. Agents Anywhere (Anthropic, OpenClaw, OpenAI)
Finally, other players have progressed by moving beyond runtime-configuration choices. Anthropic is a striking example: alongside its flagship models, it has introduced cognitive architectures that could either "shift" or accommodate the runtime layer, wherever it resides. While the advent of MCP was certainly a pivotal moment for industry standards, Claude Code's innovation—its novel use of the filesystem and local-execution sandboxing—will likely stand as an equally significant milestone, while the industry now draws broad inspiration from it (e.g. OpenAI's Codex App). Their engineering stance seems to favor no particular infrastructure location for the runtime, yet it elegantly accommodates all options.
It may be the very same infrastructure-agnostic approach—together with extreme decentralization—that propelled the viral phenomenon OpenClaw into the AI sky like a meteor, catapulting it from a weekend project by a lone veteran developer to the decade's most opportunistic acquisition in just a few weeks.
By the way, OpenAI, although deeply intertwined with Microsoft's cloud infrastructure, also occupies a distinct position. By embedding execution capabilities directly within its models and tooling—Code Interpreter, structured tool use, multimodal action orchestration—it blurs the boundary between model and runtime. Rather than owning a specific infrastructure layer, it increasingly shapes the abstraction layer above them all.
Conclusion
Once again, it's too early to predict which agent-runtime architecture will ultimately prevail—and it's likely that several will coexist. In the meantime, let's acknowledge the breakneck pace at which strategies of key AI players and global balances are retaking shape. While their public narratives often sound quite similar, the real contest is happening behind the scenes—an intense strategic race to own the execution layer that will define the next architecture of intelligence.
Meanwhile, an analogy comes to mind: millions of years ago, nature has, once again, engineered a highly versatile cognitive structure: the octopus.
It is not a fully centralized organism—though it does possess a central brain—nor is it merely a collection of independent units. In fact, most of its neurons reside in its eight arms, which by the way communicate quite independently of the brain, as evidenced by a recent study.
This is a perfect example of coordinated, distributed intelligence. Its architecture could inspire reliable systems that combine central planning cognition together with locally-empowered execution, distributed at the edge.
