The breaking change you missed

Hey there 👋

Yesterday I spent twenty minutes perfecting a prompt, then watched the whole agent fall over on a tool call. The model was fine. A session header wasn't. I'd been staring up at the smart part while the boring part quietly broke.

That's the whole week in one bug. Nobody shipped a frontier model that changes your Monday. The real moves happened one layer down — in the plumbing your agents actually run on. The protocol they speak is getting its biggest change ever. The data they run on is being rebuilt. A 200-year-old data firm put real agents into production and shipped them through that same protocol. And the training set for the next wave of real-world agents turns out to be video game clips.

The lesson keeps repeating: the leverage and the landmines both live under the model. Five minutes — let's look down.

The Big Thing

Your agents' protocol is going stateless — and the breaking change lands July 28

Here's a date for your calendar that almost nobody has: July 28. That's when the next version of the Model Context Protocol — the thing nearly every agent uses to call tools — becomes final. And it is the largest revision of MCP since launch, with breaking changes underneath it.

First, the honest framing. This is not a fresh drop. The release candidate locked on May 21, while everyone was watching model launches. So treat this as a deadline you missed the memo on, not breaking news. The clock is the story.

What actually changes: MCP is going stateless at the protocol layer. The initialize handshake is gone. The Mcp-Session-Id header that pinned a client to one server instance is gone. In plain terms, any request can now land on any server instance. A remote MCP server that used to need sticky sessions and a shared session store can now sit behind a plain round-robin load balancer and just work. That's a real operational win — horizontal scaling without the session glue.

There's more in the box. Tasks — the way servers handle long-running work — graduates from an experimental core feature to a standalone extension, so anyone who built against the old experimental Tasks API has to migrate. Three older features — Roots, Sampling, and Logging — are deprecated. Don't panic on those: the deprecations are annotation-only, and they keep working for at least twelve months. Authorization also gets hardened to line up with how OAuth and OpenID Connect actually get deployed.

So what do you do with four weeks of runway? Two paths. If you run a remote MCP server, or you shipped against the experimental Tasks API, this is a migration audit — open the draft spec, check the changelog against your transport and your task lifecycle, and validate now while it's mechanical. If you only consume MCP through a Tier-1 SDK, you mostly wait for the SDK update; those are expected to ship support inside the validation window. Either way, the move this month is the same: put July 28 on the calendar before it puts itself there.

Ship it? ACT this month — but audit, don't panic. The stateless rework is a genuine upgrade for anyone scaling agents horizontally, and the breaking parts are finite and documented. Read your own stack against the draft spec now. The teams that get burned in August will be the ones who found out about the deadline in August.

Sources: Model Context Protocol · The New Stack

Tour de Headlines

📊 Palantir and Zeta signed a 7-year deal to rebuild marketing's data layer for agents. On June 23, Palantir and Zeta Global announced a seven-year partnership to re-architect Zeta's Data Cloud on Palantir Foundry, with Zeta's Athena engine positioned as the "operating system for agentic marketing." The deal targets Palantir's enterprise base, management floated a potential $100M+ a year to Zeta over time, and ZETA stock popped on the news. Read past the number. The real signal is an ad-tech company deciding its competitive moat is now a data-infrastructure decision, not a creative one. Agentic marketing's bottleneck was never a smarter model — it's governed, connected data the agent can act on. You rebuild the pipes first; the agents are worth something after. One caveat: this is a roadmap and a go-to-market, not a feature you flip on today, and that revenue figure is Zeta's own projection.

Sources: Zeta Global · BusinessWire

🛡️ Dun & Bradstreet put compliance agents into production — and shipped them through an MCP server. On June 18, the 200-year-old data firm D&B added agentic capabilities to its Risk Analytics platform — available, in their words, "fully agentically via its MCP server," and embeddable straight into Anthropic's Claude (Copilot and Google ecosystems are next). This is the Big Thing made concrete: a regulated incumbent didn't bolt on a chatbot, it exposed real compliance workflows over MCP. That's exactly why the stateless cutover isn't a dev detail — the protocol is now the rails a compliance product rides. The early numbers are loud: processing time down 70–96%, false positives down 50–90%, review capacity up to 20x. Keep the honesty on, though — those are company-stated early results, not independent benchmarks. The durable lesson is the moat: it's the verified D&B data the agent calls, not the agent.

Sources: PR Newswire · Finextra

🎮 $320M says the next real-world agents will be trained on gameplay clips. On June 25, General Intuition raised a $320M Series A at a $2.3B valuation — led by Khosla Ventures, with General Catalyst, Eric Schmidt, and Jeff Bezos along for the ride. To teach an agent to move through space and time, you need a huge library of actions and their consequences. The cheapest one on Earth turns out to be gamers' recorded inputs — hundreds of millions of hours of gameplay, labeled frame by frame with exactly which buttons got pressed and when. The contrarian read: the next data moat is behavioral, not textual, and the people sitting on labeled human action just got interesting. (Yes, Khosla also backed last week's governance raise — the man is buying the whole agent aisle.) No product to deploy here; the money funds pre-training the next model.

Sources: TechCrunch · Dealroom

Tool of the Day

🛠️ Amazon Quick autonomous agents

Always-on, no-code agents that live in your work apps and act on a leash you set — from step-by-step approval to full goal-based autonomy.

If your stack lives on AWS, this is the week's most try-able thing. At the AWS Summit New York keynote on June 17, Amazon Quick added autonomous agents that run continuously in the background — each with its own expertise, tone, and tool access — and a dial that runs from "ask me before every step" to "here's the goal, go." No code required. They plug into CRM, email, Slack, and calendar and do real work: chasing stalled deals, processing purchase orders, summarizing regulatory changes. The builder note, said plainly: this shipped June 17, so it's not hot off the press, and it's an AWS-ecosystem play — but it's available to try now. Start the autonomy dial low, watch what the agent actually does on your real workflows, then widen the leash once it's earned it. That's the whole game with always-on agents: trust is something you grant in increments, not all at once.

See Amazon Quick agents →

Worth a Click

• xAI's Grok 4.3 is now callable in Amazon Bedrock (AWS)
  As of mid-June, Grok 4.3 runs in Amazon Bedrock — a 1M-token context window, configurable reasoning effort, strong tool use, and $1.25 per million input tokens. If you build agents on Bedrock, there's a new frontier-class model to route to.
• NVIDIA's open agent stack keeps eating the enterprise (NVIDIA)
  The open-model base layer for agents is consolidating fast. Names like CrowdStrike, Palantir, ServiceNow, Siemens, SAP, and Red Hat are building on NVIDIA's Nemotron 3 models and Agent Toolkit. Not a fresh launch (it dates to late May) — an adoption curve worth watching if you're picking a base layer.

Look down, not up. This week the action was never at the model layer — it was in the plumbing. The protocol your agents speak is changing (mark July 28). The data they run on is being rebuilt (Zeta on Foundry). Even the training corpus moved — to gameplay. The model is the part that swaps out from under you, and that's fine; it always will. The transport, the data, and the deploy surface are the parts you have to maintain. So go read your own stack today — before the calendar reads it for 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.

You're receiving this because you subscribed to The Agent Stack. · Unsubscribe