| Hey there 👋 Funny week. Nobody shipped a model that matters. The headlines you'd expect — a new frontier model, a smarter assistant — never came. Instead, the money and the launches went to the boring parts. The parts that decide whether your agents are even allowed to run. Two of those parts kept showing up. The brake — govern the agent, scope it, make it ask permission before it does something it can't undo. And the meter — bill it, measure it, watch what every call actually costs. Investors poured $30M into one company building the brake. GitHub flipped every Copilot plan to a meter. OpenAI unveiled its own inference chip — a meter built into the silicon. A healthcare agent crossed a billion-dollar valuation by owning its protocol data instead of its model. The model layer sat quiet while the control layer got loud. Underneath it all is one hard number: 97% of enterprises have deployed an agent, but only about 11% run one in production. The bottleneck moved. It's no longer "can the agent do the task?" It's "can you afford it, audit it, and trust it in prod?" The model is the part you rent. The brake and the meter are the parts you own. Five minutes — let's go.
The Big Thing Vinod Khosla Wanted Every Dollar: Runlayer Raises $30M to Govern the Agent Workforce On June 24, Runlayer announced a $30M Series A led by Felicis, with Khosla Ventures participating. Felicis preempted the round, and total raised now sits at $42M. The company came out of stealth about seven months ago on an $11M seed from the same two investors. Vinod Khosla reportedly wanted to "buy every available dollar of the round." That's the story everyone will tell: another AI governance startup raised money. Here's the part worth your attention. Founders pitch agent governance as defense — stop the rogue agent before it deletes the database. But the smart money is buying it as offense. The control layer is the thing that unlocks spend. It's what lets a CFO or CISO say yes to letting agents run at all. The brake isn't the thing that slows you down. The brake is what lets you floor the gas, because nobody hits the highway in a car with no way to stop. That framing flip is what the round exposes. An $11M seed became a preempted $30M Series A in seven months — not because governance got more interesting, but because buyers got more nervous. Every team that wants an agent in production hits the same wall: someone in security or finance has to sign their name under it. Runlayer sells the thing that makes that signature possible. What does Runlayer actually sell? Security and compliance infrastructure built around the Model Context Protocol (MCP). You see what your agents access. You control it. You require human sign-off on sensitive actions, and you keep an audit and spend trail behind it all. Named customers include Instacart and Gusto. For market context, Gartner expects 40% of enterprise apps to include AI agents by the end of 2026, up from under 5% in 2025 — so the buyers are coming whether the guardrails exist or not. Now the honest part. This is a funding round for an early company, not a product you can deploy this Friday. You can't buy Runlayer this week. So don't let the verdict be "go buy Runlayer." Let it be the principle the round is paying for. Because you can ship that principle today. Give every agent its own scoped identity — not a shared API key that does everything. Put a human gate in front of anything sensitive or irreversible. Turn on access and spend visibility before you widen the blast radius, not after the incident write-up. None of that requires a Series A. It requires a decision. Make it concrete this week. Take one agent that already runs and ask the boring questions: Whose key is it using? What can it touch that it shouldn't? What happens if it loops and burns tokens at 3 a.m.? Then fix those three — a scoped credential, a deny-list for the irreversible stuff, a spend cap with an alert. That's a half-day, not a roadmap. The teams that scale agents in 2026 won't be the ones with the best model. They'll be the ones who built the brake first, so they were allowed to press the gas at all. Ship it? WATCH the company, ACT on the principle this week. You can't deploy Runlayer today, but you can do what it sells: give every agent its own scoped identity (not a shared key), put a human gate in front of sensitive or irreversible actions, and turn on access + spend visibility before you widen the blast radius. Governance isn't the tax on shipping agents — it's the permission slip. Sources: Fortune · Runlayer · Dealroom
Tour de Headlines ⚙️ OpenAI built a chip — and used its own models to design it. On June 24, OpenAI and Broadcom unveiled "Jalapeño," OpenAI's first custom "Intelligence Processor," built for LLM inference. It's the first chip in a multi-generation OpenAI/Broadcom platform. The detail that should stick: design-to-tape-out took about nine months, and OpenAI used its own models to compress the cycle. AI is now shortening the hardware roadmap that feeds AI. The deployable read isn't "buy a Jalapeño" — you can't. It's that the platform you build on is moving to control its own unit economics, and that eventually reaches your API bill. Inference cost is the silent tax on every agent you run. Keep the hype off, though: the chip is still in testing, "performance per watt substantially better than current state-of-the-art" is OpenAI's own claim and unbenchmarked, and deployment is targeted for late 2026. Sources: OpenAI · Broadcom · TechCrunch 🏥 A healthcare agent just hit $1.2B — and the moat is the protocol data, not the model. On June 24, Assort Health raised a $120M Series C led by Menlo Ventures at a $1.2B valuation, bringing total raised past $222M. Read past the raise. This is a governed agent running at production scale in a regulated vertical — exactly the proof point your team gets asked to justify. The defensibility isn't the voice agent anyone can build. It's the data underneath: a proprietary model, "Synapse," trained on 190M patient voice interactions, 62,000 care protocols, and 1.6M decision pathways. Assort calls it the "largest deployment of AI agents for the patient journey." The lesson for any regulated vertical — own the protocol and edge-case data, because that's the asset a competitor can't clone with a better base model. One honesty note: outcome figures like "revenue up 20x in 15 months" are company-stated, not independently verified. Sources: PR Newswire · Fierce Healthcare 📊 97% deployed an agent. About 11% actually run one. Here's the gap your readers live in, as one stat. WRITER's 2026 Enterprise AI survey (n=2,400) found 97% of enterprises deployed an AI agent in the past year. McKinsey's 2026 read found only about 11% run one at genuine production scale — which leaves roughly 88% stuck in pilot purgatory: built, demoed, funded, never shipped to prod. The honest diagnosis: that gap is a governance, eval, and infrastructure problem, not a model problem. The models are good enough. What's missing is the layer that lets you trust one in production — which is exactly why the day's money is buying the control layer. The agent that ships isn't the smartest one. It's the one someone signed off on. Sources: WRITER 2026 Enterprise AI survey · McKinsey
| Sponsor You're learning to meter your agents. Who's measuring your humans? RapportScore reads the human communication signals in every sales and customer call and scores how well your team actually connects — deterministic measurement, not vibes. You're about to govern, meter, and audit every agent you deploy. Hold the conversations that close deals to the same standard. See where rapport breaks before the deal does. See your team's score → |
Tool of the Day 🛠️ Salesforce MuleSoft Agent Fabric A control plane for agents you didn't build — token/cost/data-flow visibility, an MCP bridge, and per-agent identity with mobile sign-off for high-risk actions. To set expectations: most of Agent Fabric is GA, not new — its Agent Governance features (AI Gateway, MCP Bridge, and Trusted Agent Identity with mobile authorization for high-risk actions) have rolled out over months. So treat this as a control plane to evaluate, not a launch to chase. The fresh hook is the AI Gateway's new LLM Governance — centralized visibility into token usage, cost, and data flows for third-party models, the agents you didn't build. That's the meter, applied to other people's agents. The sharpest builder takeaway, though, is "Guided Determinism" in Agent Script: it fixes the handoff rules between steps and lets the LLM reason only inside each step. Read that for what it is — an explicit concession that fully-autonomous multi-agent orchestration isn't enterprise-ready yet. So you script the path and let the model think within the rails. If you're running agents you didn't write, this is the kind of perimeter to measure against. See Agent Fabric →
Worth a Click - The meter is here: GitHub Copilot moves every plan to usage-based AI Credits (GitHub Blog)
As of June 1, every Copilot plan bills on usage — 1 AI Credit = $0.01, charged on tokens in, out, and cached. Agent mode fires multiple model calls per task, and some devs reported 10x–50x cost jumps. Price agents, not seats. - The regulatory clock — and why per-agent identity is becoming a compliance artifact (White House · Holland & Knight)
A US executive order (June 2) tells the Attorney General to prioritize prosecuting people who use agents to access data unlawfully. And on Aug 2, 2026, the EU AI Act activates transparency rules, GPAI penalty powers, and market-surveillance authority — the big high-risk Annex III obligations were pushed to Dec 2, 2027. Your audit log and per-agent identity are quietly turning into legal evidence.
The meter and the brake. This week nobody shipped a model that matters — they shipped the parts that decide whether your agents are allowed to run. The brake: govern, scope, approve (Runlayer, Agent Fabric). The meter: bill and measure (GitHub credits, the per-watt math on your own inference chip). The bottleneck stopped being "can the agent do the task?" and became "can you afford it, audit it, and trust it in production?" That's the whole 97-versus-11 gap in one sentence. The model is the part you rent — it'll swap out from under you, and that's fine. The brake and the meter are the parts you own. So build them on purpose, before the buyer, the auditor, or the bill makes you. Stay sharp — The Agent Stack
Your daily 5-minute brief on AI agents, agentic workflows, and the automation tools B2B builders actually ship. Published weekday mornings by Pixiu Media Holdings LLC. |
