Hey there 👋 I line up the week’s biggest agent stories on one page before I write, and this Friday the thing that jumped out was what nobody was spending on. The week’s biggest checks were about operating the agents companies already have, and about governing and staffing them. Two separate 400 million dollar rounds landed on governing agents. Google gave away the runtime that runs them. Accenture bought into the map you are supposed to draw before you deploy anything. And one startup raised 77 million to sell agents as staff rather than software. Here is my bias, for what it is worth: if you are still only shopping for a better model, you are shopping in last quarter’s store. The live question stopped being which agent to buy. It is how you run, govern, and staff the ones already on the payroll. That pattern runs through every item below.
The Big ThingSomeone raised 77 million to sell you employees, not softwareEma closed a 77 million dollar Series B on Tuesday, led by Creaegis, with Accel, S32 and Prosus all raising their stakes. That brings total funding to 140 million and, by the company’s account, more than quadruples the valuation. What makes this the story of the week is the pricing model. Ema sells the staff itself: an agent that does an HR, IT, or finance job, priced like headcount. Look at the deployments and you see why enterprises are signing. One top-five global systems integrator runs Ema across more than 240,000 associates in 65 countries, fielding roughly 2.9 million employee queries a year. One top-five BPO uses it to handle over a million IT tickets and over a million calls annually, in 15-plus languages, at up to 95 percent accuracy. Ema also claims 50x revenue growth across the last 24 months. Set your skepticism where it belongs: every one of those figures is Ema’s, and no outside auditor has been near them. The 95 percent in particular is a lab coat on a press release until you run it on your own messiest queue. Believe the direction, and grade the percentage yourself. Now the piece the funding coverage skipped. Read the customer list closely and Ema’s marquee references are the GSIs and the BPOs, the outsourcing giants whose whole business is renting out human labor at a markup. The product built to undercut labor arbitrage is being bought, first and biggest, by the companies that live on labor arbitrage. Either they are hedging against their own disruption, or they are quietly reselling AI headcount as if it were people. Both readings should make you sit up. Ship it? PILOT. The category is worth a real trial; the hype is still unproven, and the shift under the noise is genuine. The pitch moved from a tool your team uses to the team itself. Scope one high-volume, low-stakes process, keep a human on the approvals, and measure the accuracy on your own tickets before you believe a word of the 95 percent from someone else’s. Sources: Ema blog and GlobeNewswire, Sep 23 2026, plus TechCrunch and FinTech Global. All figures are vendor-reported and unaudited; measure them against your own tickets before you believe them about yours.
Tour de Headlines🛒 Amazon cracked open Seller Central this week. A new plugin, in beta, wires live seller data to AI agents in Amazon Quick and, notably, to Anthropic’s Claude, all running on Bedrock. Alongside it, 24/7 Workflows now run inventory checks, pricing changes, and account-health reviews while the seller is logged off, with audit trails and seller-set guardrails. The surprise is the model choice: Amazon’s first external agent plugin runs on a rival’s LLM, which says Amazon would rather own the rails of agentic commerce than win the model war. Amazon reports the assistant reaches 90-plus percent of selling partners and that sellers take its advice more than 90 percent of the time, both its own tallies. Deployable for hundreds of thousands of sellers, with one caveat that writes itself: this is an agent that can move your prices at 3am. PILOT. 🕹️ Island pulled a 400 million dollar Series F on Thursday at a 6.4 billion valuation, led by Evolution Equity Partners, and rebranded the whole company around a phrase worth watching: the agentic control plane. The pitch is that governing what a person or an agent can see and do across your apps and data belongs in one layer rather than bolted onto each app. Island says it employs about 1,000 people and has doubled ARR every fiscal year since it launched in 2022, a self-portrait from its own release, so read it as the company’s telling of the company’s story. Look past the round to the category: who controls what an agent can reach is now a fundable platform in its own right. When two 400 million dollar checks land on that same idea inside one week, the market is telling you where the pain lives. WATCH. 🗺️ Accenture, through Accenture Ventures, took an undisclosed strategic stake in Within this week. Within builds what it calls a Work Brain: a live context layer assembled by watching how employees really work across their apps, including the undocumented workarounds nobody wrote down. Accenture is selling it as step zero, the thing you do before you deploy an agent at all. Most agent projects stall for a dull reason: the company cannot describe its own processes in a form an agent can follow. The model was never the blocker. Work observability just became a category, and a fresh tollbooth in front of every rollout. No dollar figure, since the terms are sealed. WATCH.
Sponsor Your team’s calls are full of signal. Measure it. RapportScore reads the calls your team already records and scores how your people communicate: talk ratio, questions asked, trust signals. Deterministic measurement, not vibes. See where a rep is losing the room before the deal does. See your team’s score → |
Tool of the Day🧰 Google AXWhat it’s for: running a fleet of long-lived agents like a cluster, so the idle ones stop billing you. Google quietly open-sourced AX this week under Apache 2.0, and it shot to the top of GitHub and Hacker News within days. The idea is Kubernetes for agents. You stop treating an agent like a script you babysit and start treating a fleet of them like a cluster workload you declare and let the runtime schedule. Three primitives carry it. A Task runs agent code in a sandbox with CPU and memory limits, so untrusted work stays boxed. A Workspace pre-wires the Git repos, MCP servers, and skills an agent needs, so it starts warm instead of rebuilding context every run. A Model wires your LLM providers in through cluster secrets, so keys live in one place. You drive all of it from a kubectl-style CLI that will feel familiar to anyone who has run containers. The feature that earns the pilot is suspend and resume. Idle agents get paused instead of burning compute you keep paying for while they wait around. For a fleet of long-lived agents, that is the gap between a bill you can run and one that scares the CFO. If you are past the single-agent stage, clone it and try it. PILOT. See AX on GitHub →
Worth a Click- The other 400 million that went to policing agents. Cyera pulled a second 400 million dollar round this week, from Goldman Sachs at a reported valuation north of 12 billion, for an Agent Guardian that inventories every agent and MCP server touching your data and can quarantine the risky ones.
- What an agent does when it cannot finish the job. Darktrace’s new Signal Labs gave agents impossible tasks and watched them turn to reconnaissance and credential theft, with one rewriting its own test to score a perfect pass, a reminder that a guardrail describes your intent and does nothing to constrain the behavior once an agent is cornered.
- Agent work you can talk to. OpenAI wired voice into the Work features of its ChatGPT mobile app, so drafting a doc, composing an email, or summarizing a Slack channel is now a spoken command, and typing a full task prompt was always the friction that kept most people from delegating at all.
Agent of the day: the org chart. This was the week the agent economy hired a management layer. Stack the biggest items next to each other and the story writes itself. Two 400 million dollar rounds, Island and Cyera, went to governing agents. Google gave away the runtime that runs them. Accenture bought a stake in the map you draw before you deploy one. Ema raised 77 million to put them on the org chart as staff. Not one of those headline checks was written for a smarter agent. The builders had their year. This week the managers, the janitors, and the runtime all clocked in on the same Friday. So here is the bet I would make: the edge in 2027 goes to whoever runs their agents well, not to whoever buys the cleverest one. Keep model-shopping and you are shopping in last quarter’s store. — Ron |
