The dawn of the agentic era: who gets paid when software starts acting

Disclaimer: Not investment advice. I may hold positions in some of the companies mentioned. Company figures are from filings and earnings calls through September 2026; prices are closes to 25 September.

A Ghibli moment for agents

On 8 September Meta launched Muse, a personal AI agent that sends emails, books travel, fills in forms, shops and negotiates on your behalf. It keeps working after you close the app and asks for approval before anything sensitive. It runs on Meta's own Muse Spark model, pays through Stripe Link, and costs nothing to try, with $20 and $100 monthly plans for heavier use. According to Sensor Tower data reported by Bloomberg, it was downloaded 902,000 times in six days and went to number one in both US app stores, ahead of Meta's own Meta AI app over the same window. Meta's stock rose 11% on the day that report came out. At Connect two weeks later, Meta added a video avatar for the agent, control of a Mac, its own email address, a version for its glasses, and a long list of commerce partners from Shopify and PayPal to Walmart and Expedia.

Citrini Research called it agentic AI's "Studio Ghibli moment". The reference is to late March 2025, when OpenAI put image generation into ChatGPT and the internet turned itself into Hayao Miyazaki drawings. Within two days Sam Altman was writing that OpenAI's GPUs were melting. By the following week ChatGPT had added a million users in an hour, and 130 million people had made more than 700 million images. A capability that power users had had for a while got a wrapper ordinary people wanted, and demand broke the infrastructure behind it.

The analogy holds in one important way: agents were already here for developers, and Muse is the first time they've been handed to the general public in a form people seem to want. It breaks down in two others.

First, a Ghibli portrait is harmless. An agent spends your money, speaks in your name and has your passwords. One poll reported this month found only 7% of people would let an agent buy things without asking first. Amazon blocked Muse from its store at launch, citing permission and identification concerns.

Second, not all of Muse is yet what it looks like. 404 Media reported that some of the phone calls Muse placed to restaurants and shops "on behalf of" users were made by human contractors in call centres, and that some users weren't told. Follow-up reports say the test has been rolled back, and Meta says it will launch the feature only with proper disclosure. Tellingly, the human-completed calls reportedly succeeded far more often than the AI-only ones.

So the consumer moment is real, and it's rough. That's usually how these things start. The part I find more investable is less visible: what an agent actually does when it runs, and who gets paid for each step.

What an agent is made of

An agent is a language model plus a harness. The model does the reasoning. The harness is everything that lets that reasoning touch the world: tools it can call, a workspace with files and code, memory that persists between steps, a network to run on, and guardrails and logs so someone can see what it did. Meta runs every Muse agent inside its own small virtual machine, two processor cores and eight gigabytes of memory, which is a reminder that the harness is physical as well as conceptual.

Fig. 1What an agent is made of

The distinction matters economically. As the models converge and get cheaper (the subject of my piece on the stacked S-curves), the scarce thing is less the reasoning and more the harness around it. And a harness has a property the model doesn't: nearly everything it does is metered by someone other than the model provider.

The demand is already here

The consumer story is a few weeks old. The enterprise story is more than a year old, and it's large.

Fig. 2Somebody is already paying for agents

Anthropic's annualised revenue was about $1 billion as 2024 closed; by late July 2026 it had passed $65 billion. Its CEO has said the company planned for tenfold annual growth and got eightyfold. Much of it comes from coding, and most of it arrives through the API from companies building their own products and agents. OpenAI crossed $40 billion in August. However you feel about the valuations of these labs, those are customers paying for agents to do work.

Fig. 3Tokens, then more tokens

The token data tells the same story from another angle. OpenRouter, which routes developer traffic across dozens of model providers, went from half a trillion tokens a week in January 2025 to about 126 trillion by mid-September 2026. It says early February 2026 was probably the last day humans consumed more tokens on its platform than agents did. Cloudflare says non-human traffic overtook human traffic across its network in May 2026, well ahead of its CEO's own forecasts. The Model Context Protocol, the standard that lets agents plug into tools, is downloaded more than 97 million times a month.

Fig. 4Why agents are a multiplier

Why agents multiply demand rather than just add to it: a chat is one question and one answer. An agent plans, calls a tool, reads the result, calls another, checks its work and tries again. Anthropic measured its own agents using about four times the tokens of a chat, and multi-agent systems about fifteen times. Datadog's CEO made the point that matters for everyone else: agents spend much of their time, sometimes most of it, calling tools, and those tools are ordinary software running on ordinary servers. Every one of those calls is billable by somebody.

Where the meter runs

That gives me a simple test for who benefits: follow an agent through a task and list what it touches. It usually needs to reach a person, write or change code, move across the network, and be monitored while it does. Four public companies sit on those four steps and charge by usage rather than by human seat.

Fig. 5Three accelerating, one not yet

The chart is the scorecard. Three of the four have turned from slowing growth to accelerating growth over the last year and a half. The fourth, GitLab, shows the pattern before it reaches reported revenue. Here is what each one does for an agent, and the evidence that agents are driving it.

Twilio: agents need to reach people

Agents that book, sell, support and remind need to reach humans, and humans still mostly answer phones and read texts. Twilio sells the plumbing for that. Its new conversations layer, generally available since May, connects a phone call to an AI model and handles the speech both ways, with memory and analysis on top, all metered.

Fig. 6A minute an agent talks is worth more

An ordinary inbound voice minute lists at under a cent. The same minute with ConversationRelay attached lists at nearly eight cents, before memory and analysis charges. The company's numbers are starting to reflect it. Voice revenue grew more than 20% last quarter and self-serve voice more than 50%, and non-GAAP gross profit growth has accelerated for five quarters in a row, to 18%. Gross profit is the cleanest measure for Twilio, because pass-through carrier fees inflate revenue but add no profit. Management described one AI-native customer going from low-six-figure quarterly spend to a $6 million annual run-rate in a little over a year. Twilio has raised its organic growth guidance for the year twice, from 8–9% to 13–13.5%.

A caution, because the stock moved on it: Twilio's link to Muse is unconfirmed. The shares jumped about 23% in four sessions around Meta Connect as the market cast Twilio as Muse's phone company. Neither company has said so. The one first-hand account from a developer testing Muse's calling says Meta handles telephony itself. HSBC downgraded the stock on the 25th, arguing that Meta could connect directly to carriers and that even a very bullish Muse scenario adds under 1% to Twilio's revenue. Bandwidth, which nobody claims has a Muse contract, rose almost exactly as much that week, which tells you it was a theme trade. The Twilio story I find convincing is enterprises building their own voice agents on its platform, not one consumer app.

Datadog: somebody has to watch the agents

Traditional software does the same thing every time. Agents don't. They take different paths through the same task, call different tools, and fail in ways nobody predicted. That makes monitoring them (tracing each step, measuring cost and latency, catching loops and bad tool calls) far more important than monitoring a conventional app.

Datadog is where that data lands. Revenue growth has accelerated for five straight quarters to 36%. It has more than 750 AI-native customers, eight of them spending over $10 million a year. Calls from agents to Datadog's own tool endpoints grew fourfold in a single quarter and more than 22-fold since the end of 2025. Crucially, growth outside the AI-native companies has accelerated too, from 18% a year earlier to the high twenties, which is the evidence I care about: ordinary enterprises deploying agents and needing to watch them.

The risk is concentration. Datadog's largest customer, widely reported to be OpenAI although Datadog doesn't name it, renewed with reduced usage, and the stock fell 19% on a quarter that beat estimates. Management says growth excluding that customer is roughly the same. I take the market's reaction as a reminder that the biggest AI-native customers can and will build their own tools; the broader enterprise base is what makes this durable.

GitLab: agents write code, and code needs a home

Coding is the first killer application of agents, and every line an agent writes has to be stored, versioned, tested, scanned and approved before it ships. GitHub's own numbers show the scale: monthly commits on GitHub went from 1.4 billion in April to 2.9 billion in August 2026. On 17 August GitHub suffered a global outage of almost eight hours, which its CTO described as a capacity failure at its core.

GitLab is the other major home for code and the one enterprises choose when they want to run it themselves or stay independent of any single cloud. It's also the most interesting of the four, because its reported numbers and its underlying numbers are telling different stories.

Fig. 7The load arrives before the revenue

Reported revenue growth has slowed every quarter for two years, to 21%. That's why I listed it as a counterexample in the SaaSpocalypse piece. Look one layer down, though, and everything is accelerating. Code pushes are up 50% year on year, pipelines 40%, repositories secured 60%. First orders from new customers doubled, and the net new recurring revenue signed in the quarter grew 42%. Management says heavy AI adopters have seen codebases grow as much as fivefold, and it is rebuilding Git itself, with Anthropic as design partner, for roughly a hundred times human scale.

The monetisation is only just starting. GitLab's usage-priced agent platform went from $15 million to more than $40 million of annualised paid consumption in one quarter, with a target of over $100 million by January. And about 80% of queries to its new code-context service come from outside agents such as Claude Code and Codex. GitLab gets paid when agents it didn't build do work on code it hosts. That's the harness thesis in one data point.

The risks are real. Gross margin has fallen from 90% to 86% as it pays for AI inference. Some of the new consumption revenue may be repackaged seat money, which management says is too early to split out. And GitHub, with Microsoft's distribution and millions of paid Copilot users, is fixing its capacity problems.

Cloudflare: the network agents run on

Cloudflare sits between much of the internet and its users, which makes it the place to watch what agents are doing to traffic. It reported that non-human traffic passed human traffic on its network in May. AI-training crawlers went from 22% to 52% of crawler requests in a year, and in categories that are heavily crawled, human visits have fallen by as much as 40%. Its response has been to charge: pay-per-crawl for AI companies since 2025, per-use pricing since July, and default blocking of mixed-use crawlers on ad-supported pages since 15 September.

It also runs agents directly. Its Workers platform, which runs code at the edge, has become a meaningful contributor to revenue and the fastest-growing product line. Recent deals include multimillion-dollar commitments specifically to host secure containers for AI agents, and the popular vibe-coding platforms deploy their users' apps on it. Its CEO's argument is about scarcity: if every knowledge worker ran an agent in its own container, the world wouldn't have enough CPUs, and Cloudflare's lighter-weight approach uses far fewer. Revenue growth has re-accelerated from 26% to 36%, customers paying over $100,000 a year are up 27%, and it added about 2 million developers in the last quarter, more than in all of 2025. The agent-payments business it's building is still pre-revenue, and much of the bot traffic it handles, it blocks for free.

The market has noticed

Fig. 8The market has noticed

None of this is a secret. Datadog, Twilio and Cloudflare are up between 77% and 97% this year, while the software ETF is flat. Cloudflare trades at about 33 times next year's sales and more than 200 times next year's earnings. Datadog is at about 17 times sales. Those prices leave little room for a stumble like Datadog's August reaction. GitLab and Twilio are the least demanding of the four on revenue multiples, and GitLab is the only one whose acceleration isn't yet visible in its reported revenue.

What could go wrong

• Agents disappoint. Muse's human call centre is the example to remember. If agents turn out less reliable than hoped, usage grows slower. Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027.
• The biggest customers build their own. OpenAI reducing its Datadog usage and Meta potentially going directly to carriers are the same risk: the largest AI companies have the scale to bring the harness in-house. The defence is the long tail of ordinary enterprises that won't.
• The labs bundle the harness. A model provider that offers calling, hosting, code storage and monitoring inside its own platform would squeeze all four. I'd watch for this most closely at OpenAI and Anthropic.
• Inference eats margins. GitLab's and Cloudflare's gross margins are both down year on year as they pay for AI workloads. Usage revenue at a lower margin is still good, but the multiples assume software margins.
• Price. For three of the four, the market has already paid for a lot of this.

The dawn, not the day

The agentic era looks a lot like the fourth step in the staircase I described in the first essay of this series: a capability humans built by giving models tools and time, now spreading from developers to enterprises and, with Muse, to everyone. Whether or not models ever start improving themselves, that step creates a durable pattern. The more work agents do, the more calls, commits, requests and logs flow through the harness, and the harness is mostly metered by companies that aren't model labs.

The consumer version is at its Ghibli stage: viral, rough and partly propped up by humans behind the curtain. The enterprise version is already a $100-billion-a-year run-rate at the labs, and it's showing up in the growth rates of the companies that carry agents' work. The next place I want to follow that thread is security, because a world with far more autonomous software acting on its own is, first of all, a world with a much larger attack surface.