Image by Pawel Czerwinski
Busy 48 hours in the agent infrastructure space. Five developments that, taken together, tell a clear story: the orchestration and execution layers are where the real competition is happening. Models are commoditizing. The value is in governance, observability, and owning the workflow surface.
Cursor Blows Past $2B ARR — Doubles in Three Months
Bloomberg reports Cursor's annualized revenue topped $2 billion in February, doubling from just three months prior. Enterprise customers now account for roughly 60% of revenue. The AI coding assistant has become the fastest-growing dev tool in history.
$2B ARR isn't a coding tool story — it's a distribution story. Cursor won by owning the developer's daily context window, not by having the best model. The 60% enterprise mix is the real headline: companies are buying seat licenses for AI coding, not experimenting with it. Whoever owns the workflow layer captures the value, regardless of which model runs underneath.
Google ADK Isn't a Toolkit — It's an Agent Execution Framework
Google expanded its Agent Development Kit integrations to include GitHub, GitLab, Jira, Confluence, MongoDB, Pinecone, and seven observability platforms with native OpenTelemetry support. Analysts at Futurum Group argue ADK is positioning as the execution layer where agents run in production — not just where they're built.
The framework wars just shifted terrain. LangChain, CrewAI, and AutoGen compete on orchestration in app code. Google is going after the layer below that — the execution substrate that connects agents to real engineering toolchains. Seven observability integrations at launch is the tell: Google understands that enterprises won't deploy agents they can't observe. The question is no longer "which framework do I build with" — it's "which framework owns my production stack."
Alibaba Open-Sources OpenSandbox
Alibaba released OpenSandbox under Apache 2.0, providing isolated environments for code execution, web browsing, GUI interaction, and RL training. It supports Docker locally and Kubernetes at scale, with SDKs for Python, TypeScript, and Java/Kotlin. Integrates natively with Claude Code, Gemini CLI, OpenAI Codex, LangGraph, and Google ADK. Hit 3,800+ GitHub stars in two days.
The "execution layer" of the agent stack is getting serious attention — and it should. Every team building autonomous agents has hacked together their own Docker-based sandbox. OpenSandbox is a bet that this should be a shared primitive, not bespoke infrastructure. The cross-framework integration list is notable: this isn't vendor lock-in, it's a utility play. If it works, it becomes invisible infrastructure. That's exactly where the highest-leverage open-source projects live.
Huawei Open-Sources A2A-T Protocol Stack
At MWC 2026, Huawei announced it will open-source the core software supporting A2A-T (Agent-to-Agent for Telecom), a TM Forum-backed standard for multi-agent collaboration. The release includes a Protocol SDK, Registry Center for auth/addressing, and an Orchestration Center with low-code workflow tooling.
Agent-to-agent communication standards are inevitable — the only question is who writes them. Telecom is first because multi-vendor, cross-boundary orchestration is their daily reality. But watch this pattern: every industry with complex multi-party workflows — supply chain, healthcare, finance — will need something like A2A-T. The open-source move is smart. Standards without implementations are just PDFs. This one ships with code.
NullClaw: A 678 KB Agent Framework in Zig That Boots in 2ms
NullClaw is an AI agent framework written entirely in Zig — no runtime, no garbage collector. It compiles to a 678 KB binary, runs on ~1 MB RAM, and boots in under 2 milliseconds. Supports 22+ AI providers, MCP integration, and native hardware peripheral support for Arduino, Raspberry Pi, and STM32. The codebase includes 2,738 tests across 45,000 lines of Zig.
This is a fascinating counterpoint to the "agents need cloud infrastructure" narrative. If agents are going to live on edge devices, in embedded systems, and on $5 hardware, the Python/Go runtime tax is a non-starter. NullClaw probably won't replace LangChain on your server — but it might put an agent on every sensor, every microcontroller, every edge node. The agent future isn't just cloud-scale orchestration. It's also millions of tiny agents running where the data lives.
The Pattern
Five stories, one theme. The stack is splitting into layers, and the orchestration and execution layers are where the value is concentrating. Cursor owns the workflow surface. Google is positioning ADK as the production substrate. Alibaba is commoditizing sandboxing. Huawei is standardizing agent-to-agent communication. And NullClaw is proving that the execution layer extends all the way down to $5 microcontrollers.
Build accordingly.
Happy thinking, Jason