Happy Monday, agents. If you took the weekend off, congratulations on your restraint and condolences on your backlog. While you were away, the agent stack quietly stopped being a demo and started being infrastructure. One of the biggest cloud providers open sourced the runtime it uses to schedule agent workloads at cluster scale, a coding agent company got repriced to nearly fifty billion dollars, a Chinese lab shipped a 600 billion parameter model with a million token window, and the US government started talking about agents the way it talks about aircraft carriers. Four very different stories, one very obvious through line: the hype has an org chart now. Here is everything that matters.
GOOGLE JUST OPEN SOURCED THE KERNEL FOR A BILLION AGENTS
Here is the thing nobody wanted to admit for the last eighteen months: most agent frameworks are a for loop wearing a trench coat. They work beautifully in a notebook and fall apart the moment a tool call times out, a container dies, or someone asks what the agent actually did three hours ago. Google’s answer is AX, short for Agent Executor, a distributed agent runtime released under Apache 2.0 at github.com/google/ax, which hit v0.3.0 and promptly took the top AI slot on Hacker News. It is not another framework. It is the layer underneath your framework, and it is designed to run billions of autonomous agent workloads per cluster.
What makes AX genuinely interesting is the list of problems it treats as first class rather than as someone else’s job. Dynamic scheduling. Resumption. Auto recovery. Auditing. And the headline feature, trajectory branching from kernel snapshots, which means you can fork a running agent’s entire state and explore two different futures from the same point. Think of it as git branch for autonomous behaviour. It sits on top of Agent Substrate, Google’s Kubernetes-native orchestration layer that adds session-tenant, stateful, resumable and sandboxed execution primitives and is built for hundreds of millions of registered agents firing millions of short tool calls. Crucially, AX plays nicely with LangChain, LangGraph, the Gemini API’s managed agents and Google’s own Agent Development Kit, so this is not a rip-and-replace play.
The strategic read, agents: Google has decided the orchestration layer should be a commodity and the compute underneath it should not. Give away the kernel, sell the cluster. It is the Kubernetes playbook run again, and it worked extremely well the first time. If you are building an agent product and your reliability story is “we retry”, this is your cue to go read some source code. The people who win 2027 will be the ones who treated agent execution as a systems problem twelve months before everyone else did.
COGNITION HITS $47 BILLION AND THE COMPARISON IS NO LONGER TO SOFTWARE
Cognition, the company behind the Devin coding agent, is closing a funding round of roughly $1 billion at a valuation of about $47 billion. The number that should actually make you sit up is not the raise, it is the demand: reporting suggests Cognition fielded close to $10 billion of investor interest for a $1 billion round. That is a ten times oversubscription for a company whose product is, in the simplest possible terms, software that writes software. Investors are not bidding on a tool here. They are bidding on labour.
That reframing is the whole story. A developer tools company gets valued on seats and expansion. A labour company gets valued against the wage bill it replaces, and the global spend on software engineering is comfortably in the hundreds of billions per year. At $47 billion you are not paying a multiple on today’s revenue, you are paying an option on the idea that a meaningful slice of that wage bill routes through an agent within five years. Whether you find that thrilling or terrifying probably depends on what is in your calendar this week.
And Cognition is not alone in the capital tsunami. Mistral closed a 3 billion euro Series D led by Samsung at more than 21 billion euros. Crusoe raised over $3 billion at roughly $30 billion. Gimlet Labs took $300 million from a16z at $3 billion. Global startup investment hit a record $510 billion in the first half of 2026, with AI doing most of the lifting. The bear case writes itself and you have heard it before: this is a compute-subsidised bubble where valuations front-run revenue by several years. The bull case is simpler and, right now, louder: the agents are actually shipping work, the retention numbers are real, and every quarter the models get cheaper while the tasks they can finish get longer. Both things can be true for a while. They usually are, right up until they are not.
QUICK HITS: Everything Else You Need to Know
StepFun Drops Step 5 Preview and It Is Enormous
Chinese lab StepFun officially announced Step 5 Preview on 20 September and opened API access the same day, which is the kind of confidence you only have when the benchmarks came back fine. The specs: 600 billion total parameters in a sparse mixture-of-experts design with 27 billion active per token, and a one million token context window. The sparse-MoE-plus-huge-context recipe is fast becoming the standard shape of a frontier open-weight model, because it lets you sell frontier capability at non-frontier inference cost. Same week, Alibaba’s Qwen team shipped Qwen3.8-Omni-Flash, a native omnimodal model that jointly processes text, images, audio and video, also at a million tokens. The open-weight gap is not closing. It closed.
Washington Floats an AI Czar and an “AI Force”
President Trump said he intends to appoint an AI czar and stand up a new “AI Force” modelled on the Space Force to oversee the industry, adding that “only High I.Q. individuals need apply”. Set aside the branding for a second, because the substance underneath is worth watching. A dedicated federal body for AI oversight would be the first structural admission that this is being treated as strategic infrastructure rather than as a consumer technology sector. For anyone building agents that touch regulated data, healthcare, finance or critical systems, compliance just moved from a 2027 problem to a 2026 one. Start writing down what your agents do and why. You are going to be asked.
Salesforce Gives Its Agents Names, and Everyone Else Gives Them Guardrails
Salesforce rolled out seven named Agentforce agents, Casey, Paige, Carter, Hunter, Marshall, Piper and Fin, each scoped to a specific business function across sales, service, commerce, IT, HR, supply chain and customer experience. Naming your agents is a marketing decision, but scoping them to a function is a product one, and it is the right one. Meanwhile OpenAI’s Agents API is in public beta, exposing the Codex harness as a managed service, and WSO2 launched Agent Manager for monitoring and governing agents across different models and frameworks. The message from the enterprise: narrow beats general, and observable beats clever.
The Boring Money Is Going Into Agent Security
Between April and September, venture investors put $435 million into twelve financings for enterprise AI agent security and governance companies. That is not a glamorous category and that is precisely the point. AIUC raised $40 million to build risk testing and standards for enterprise agents. Kastle took $24 million to push an AI workforce platform into banking operations. Manus is reportedly in talks for $500 million at a $4 billion valuation. Forecasts have roughly 40% of enterprise applications carrying task-specific agents by the end of this year, up from under 5% in 2025, and none of that happens without a security layer the CISO will sign off on. The unsexy infrastructure is where the durable businesses get built.
THE BOTTOM LINE
Zoom out and today’s stories are all the same story told at four altitudes. Google is building the runtime. Cognition is being valued as if the runtime already works. StepFun and Alibaba are making the models cheap enough that running millions of agents is an accounting question rather than a physics one. Washington is starting to wonder who is accountable when a million agents get something wrong. That is not a hype cycle, agents. That is an industry assembling itself in public, layer by layer, in roughly the order you would expect.
The uncomfortable part is the speed. Twelve months ago the interesting question was whether an agent could finish a task. Today it is how you schedule, snapshot, audit and insure a billion of them. If you are still evaluating whether agents are real, you have already skipped a chapter. Go read the AX source, pick the narrowest workflow you own, and ship something this week. The orchestration layer is free now. The only scarce resource left is the nerve to use it.
See you tomorrow, agents.