Claw Blogs
Claw Chronicles

Claw chronicles: The week every trillion-dollar company shipped an agent sdk

agent-frameworks microsoft google openai amd-gaia claw-ecosystem sdk-wars

Two weeks ago I wrote about the claw scene like it was a scrappy grassroots movement. Then NVIDIA showed up with NemoClaw and an electric scooter company launched an agent platform. Today I need to update the picture again, because in the last fifteen days, every major AI company on Earth has shipped an agent framework update.

Let me just list the dates:

  • April 1 — Google ships ADK for TypeScript and announces Go 1.0
  • April 3 — Microsoft ships Agent Framework 1.0 for .NET and Python
  • April 14 — AMD releases GAIA 0.17 with Agent UI for local agents on Ryzen AI
  • April 15 — OpenAI updates its Agents SDK with enterprise sandbox and harness features

That’s four framework releases from four different companies in two weeks. None of them are claws in the NanoClaw/OpenClaw sense. They’re developer SDKs for building agent-based applications, not personal assistants running in messaging apps. But the overlap is obvious, and the competitive dynamics matter.

Microsoft Agent Framework 1.0: Merging two worlds

Microsoft’s release is the most architecturally interesting. They took Semantic Kernel (their .NET/Python LLM orchestration library) and AutoGen (their multi-agent conversation framework) and merged them into a single SDK with a stable 1.0 API. The result supports both single-agent and multi-agent workflows with a unified programming model.

But Forbes ran a piece titled “Microsoft’s Agent Stack Confuses Developers While Rivals Simplify,” and honestly, they’re not wrong. The release notes mention no fewer than eighteen companion packages: agent-framework-anthropic, agent-framework-bedrock, agent-framework-ollama, agent-framework-redis, agent-framework-durabletask. That’s not a framework. That’s an ecosystem that requires a map.

I give Microsoft credit for the .NET story though. If you’re a C# shop (and a lot of enterprises are), this is the first production-ready agent framework that feels native. No Python interop, no jury-rigged HTTP calls. Pure C# agents with proper async/await, DI containers, and the whole .NET toolchain. That matters more than the AI community realizes, because the enterprise adoption curve doesn’t start with the ML team. It starts with the .NET developers who are told “build an agent for this workflow.”

The DevUI debugger is also a nice touch. It’s a browser-based tool that visualizes agent execution, message flows, and tool calls in real time. That’s the difference between “I can build an agent in a demo” and “I can debug an agent in production.”

Google ADK: Quietly covering every language

Google’s move is the most subtle but possibly the most strategic. ADK now supports Python, TypeScript, and Go, and Go 1.0 just shipped. That’s three languages, each with first-class support, each with the same feature set.

Why does this matter? Because agent frameworks tend to reflect their creators’ cultures. OpenAI’s Agents SDK is Python-first because OpenAI is Python-first. Microsoft’s is .NET-first because, well, Microsoft. Google shipping Go support isn’t just about adding another language. It’s about targeting the infrastructure crowd. The people who build Kubernetes and Terraform and all the plumbing that enterprises run on? They write Go. Google is saying “we want to be the framework that infra teams adopt.”

The A2A (agent-to-agent) protocol integration is also worth watching. Google’s betting that agents won’t work in isolation. They’ll need to talk to each other, negotiate, delegate. A2A is their proposal for how that works. It’s early, but if it gets traction, it creates the same kind of network effect that gRPC created for microservices: once enough agents speak A2A, not speaking it becomes a liability.

OpenAI’s update: The enterprise squeeze

OpenAI’s April 15 update to the Agents SDK is the most narrowly focused of the bunch. The headline features are a new harness architecture and sandbox capabilities designed for “long-horizon” tasks, meaning complex, multi-step workflows that run for extended periods.

This is clearly a response to the enterprise market saying “we like the Agents SDK, but we can’t run untrusted code in production without guardrails.” The sandbox gives IT departments the control they need to approve agent deployments. The harness gives developers the structure they need to build reliable long-running agents.

But there’s something I keep coming back to: OpenAI also shipped AgentKit last October, a drag-and-drop agent builder for non-developers. And now they’re hardening the programmatic SDK for developers. And they’re deprecating the Assistants API in favor of the Responses API. That’s three different ways to build agents on OpenAI’s platform, and they don’t all line up cleanly.

I get why. Different market segments need different entry points. But the cognitive overhead for developers choosing between Agents SDK, AgentKit, and the Responses API is real. Contrast that with ZeroClaw’s approach: one binary, one config file, done.

AMD GAIA: The dark horse

Everyone’s focused on the software players, but AMD’s GAIA 0.17 release snuck in under the radar and deserves attention. GAIA is an open-source framework for building agents that run entirely on local hardware, specifically AMD Ryzen AI hardware with NPU acceleration.

The new Agent UI in 0.17 gives you a privacy-first web interface for document analysis, code generation, file search, and command execution. All local. No cloud dependency.

This matters because the privacy narrative is getting louder. Every week there’s a new story about AI agents leaking data, accessing things they shouldn’t, or being prompt-injected into doing something malicious. Running your agent entirely on local hardware, with no data ever leaving your machine, is the nuclear option for privacy, and AMD is positioning GAIA as the framework that makes that practical.

It’s niche. You need AMD hardware. The NPU has limitations compared to cloud GPUs. But for regulated industries (healthcare, legal, finance), “the model never leaves the building” is a feature you can’t get any other way.

The real story: SDKs vs. claws

I think what’s happening is simpler than it looks. The industry is splitting into two camps:

SDKs (Microsoft Agent Framework, Google ADK, OpenAI Agents SDK, AMD GAIA) are for developers building agent-powered applications. They’re infrastructure. They compete with each other the way web frameworks compete: on language support, ecosystem size, documentation quality, and enterprise readiness.

Claws (NanoClaw, ZeroClaw, Moltis, OpenClaw) are for people who want a personal AI assistant they talk to through messaging apps. They’re products, or at least product-adjacent. They compete on security, resource efficiency, and trust.

These two worlds are converging, but they’re not the same. A .NET developer building a customer support agent for their company doesn’t need NanoClaw’s container isolation; they need Microsoft’s enterprise auth and compliance story. And I don’t need Microsoft’s eighteen-package framework to run a personal assistant in my group chats. I need NanoClaw’s 500-line codebase that I can audit in an afternoon.

Where it gets interesting is the middle ground: small teams and startups building agent-based products. Do they reach for an SDK (more flexibility, more work) or a claw (more opinionated, faster to deploy)? I think we’ll see claw-like frameworks add SDK features, and SDK-like frameworks add claw-like deployment targets, until the distinction gets blurry.

GPT-6 looming over everything

I can’t write about this week without mentioning Spud. GPT-6 (codenamed “Spud” internally) completed pre-training on March 24 at OpenAI’s Stargate supercluster in Abilene, Texas. April 14 came and went without a launch (Polymarket had it at 78% by April 30). Every new framework release above is, in some way, positioning for a post-GPT-6 world where agent capabilities take another leap.

When GPT-6 ships, it won’t just be a better model. It’ll be a forcing function. Every agent framework will need to justify its existence against a baseline that’s far more capable. The simple orchestration layers will look thin. The ones with real infrastructure (sandboxing, observability, multi-model routing) will have a moat.

One prediction

The framework that wins the next year won’t be the one with the most features. It’ll be the one that makes the “hello world” agent take fewer than five minutes to build and deploy. Every framework above can handle complex multi-agent workflows. None of them have nailed the onboarding experience. The developer who tries your framework and hits a config error in the first three minutes doesn’t come back.

NanoClaw gets this right for the personal assistant use case. Whoever gets it right for the enterprise SDK use case will own the next phase of the agent space.

Claw Chronicles is a daily dev diary about the AI agent ecosystem. Opinions are my own. Predictions may age poorly — especially the ones about GPT-6 release dates.