MCP Drops Sessions in 2026: Why Ad Tech Must Rebuild
MCP just removed sessions, its biggest change since launch. The July 28, 2026 spec forces ad platforms to rebuild agent servers. Here is what it means for you in plain English.
📰 What Just Happened: MCP Removed Sessions Entirely
On July 28, 2026, the Model Context Protocol (MCP) shipped its largest specification update since Anthropic introduced the standard in late 2024. The new version, labeled 2026-07-28, removes protocol-level sessions completely. Six Specification Enhancement Proposals (SEPs) rewrote the transport layer so that every request an AI agent makes is now self-contained. A release candidate went out on May 21, 2026, giving server builders a ten-week window to test before the final spec landed.
If that sounds abstract, here is the short version. MCP is the plumbing that lets AI assistants like Claude, ChatGPT, and Gemini connect to outside tools: your calendar, your ad account, your e-commerce store. Until now, that plumbing worked like a phone call. The agent dialed in, a session opened, and every follow-up request had to go back to the exact same server that answered the call. The new spec works more like sending letters. Each request carries everything the server needs, so any server instance can answer any request.
The trade press picked this up quickly because the change is not optional. PPC Land reported that platforms running MCP servers, including Google Ads, Amazon Ads, Meta, Snap, and Pinterest, built their agent infrastructure around the old session model. That machinery now has to change.
🔧 Sessions vs. Handles: What Actually Changed
The old design used an initialize handshake. The client and server exchanged capabilities once, the server issued a session ID, and that ID pinned all traffic to the server instance that created it. To make this work at scale, companies needed sticky load balancer routing, shared session stores such as Redis, and gateways that inspected request bodies to figure out where traffic belonged.
The new design replaces hidden session state with explicit handles. Instead of the server quietly remembering your shopping cart, it returns an identifier like a basket_id or workflow_run_id directly to the agent as a tool result. The agent passes that handle back on the next call. State still exists, but now the AI model can see it, reason about it, and resume it later.
This matters for reliability. When state lives in a handle instead of a fragile session, an agent can survive a dropped connection, hand a task to another agent, or pick up work hours later without starting over.
The same change in one table
The comparison below shows why infrastructure teams care so much about this update.
| Aspect | Old MCP (sessions) | New MCP (2026-07-28) |
|---|---|---|
| State | Hidden inside the server session | Explicit handles the agent can see |
| Routing | Sticky sessions pinned to one server | Any instance handles any request |
| Infrastructure | Redis session stores, gateway inspection | Ordinary round-robin load balancing |
| Failure recovery | Dropped session often means starting over | Agent resumes using the handle |
| Tool discovery | Long-lived event streams | Cacheable tools/list with a TTL |
💡 Why This Matters Even If You Never Touch a Server
You might be reading this as a solopreneur who runs a few ad campaigns and uses Claude or ChatGPT for daily work. Here is why this backend change reaches you.
First, agent reliability improves. If you have ever asked an AI assistant to do a multi-step task through a connected tool and watched it fail halfway, session fragility was often part of the problem. Stateless servers with explicit handles mean agents can retry, resume, and recover without losing your work. As assistants like Claude Sonnet 4.6, GPT-4o, and Gemini 2.0 lean harder on connected tools, that stability is the difference between a demo and a dependable coworker.
Second, the cost of running an agent-accessible service just dropped. PPC Land noted that the old model demanded sticky sessions, shared session stores, and deep packet inspection at the gateway. That infrastructure priced out smaller publishers and indie tool makers. Removing it lowers the barrier, so expect more small services, including the niche tools solopreneurs actually use, to expose MCP servers.
Third, agentic advertising is coming into focus. Platforms are building MCP servers so AI agents can manage campaigns, pull reports, and adjust budgets on your behalf. The protocol going stateless is a signal that the industry expects agent traffic at real scale, not as an experiment.
🏗️ Who Has to Rebuild, and Who Gets a Head Start
The rebuild burden falls on anyone operating an MCP server at production scale. Ad tech is exposed because major platforms wired their agent endpoints to the session model early. Teams now need to move from sticky routing to standard load balancing, retire session stores where they only existed to serve MCP, and adopt the hardened OAuth authorization flows that shipped alongside the transport changes.
Some companies saw this coming. Daniel Jaye, a longtime ad-tech and martech figure, pointed out that Aqfer's cloud-native architecture had already adopted stateless, scale-out execution before the protocol made the same move. Companies built that way mostly need to update protocol handling. Companies that leaned on session machinery face a genuine migration project.
The update also reorganized the ecosystem around the core spec. Tasks moved from an experimental feature to a formally governed extension, an MCP Apps extension now covers interactive HTML interfaces, and a new extensions framework gives add-ons their own versioning. In short, MCP is maturing from a promising standard into infrastructure that large platforms plan around.
✅ What You Can Do Today
You do not need to write code to act on this news. The practical moves depend on which hat you wear.
If you use AI assistants with connected tools, expect some connectors to get flaky during the migration window and then get noticeably more reliable. If a connection breaks in the coming weeks, reconnecting the integration usually resolves it because providers are rolling out updated servers.
If you run ads, ask your platforms and tool vendors one question: are your MCP endpoints updated to the 2026-07-28 spec? Vendors who answer clearly are the ones preparing for agent-driven campaign management. Vendors who cannot answer are worth watching carefully before you let an agent touch your budget.
If you offer a SaaS product or content service, this is the moment to consider exposing an MCP server, because the stateless design makes hosting one far cheaper than it was six months ago. The official documentation at modelcontextprotocol.io covers the new spec, and Anthropic, OpenAI, and Google all publish guides for connecting tools to their assistants.
- ✔Reconnect any AI tool integrations that misbehave during the migration window
- ✔Ask ad platforms and vendors whether their MCP servers target the 2026-07-28 spec
- ✔Review which tools your AI assistant can access and remove ones you no longer use
- ✔If you run a product, read the new spec at modelcontextprotocol.io before building
- ✔Watch for platforms announcing agent-managed campaign features in the coming months
🔭 The Bigger Picture: Agents Are Becoming Infrastructure
Step back and the pattern is clear. Protocols only go through painful breaking changes when real traffic demands it. MCP launched in November 2024 as an Anthropic project, was adopted by OpenAI and Google within a year, and now carries enough production load that its session model became a bottleneck worth breaking.
For ad tech specifically, this is the plumbing being laid for a world where an AI agent, not a human in a dashboard, files the report, shifts the budget, and pauses the underperforming campaign. The platforms rebuilding their servers this quarter are not doing it for fun. They are doing it because they expect your future account manager to be an agent.
For everyone else, the takeaway is simpler. The AI tools you use are getting a sturdier foundation, and the services you rely on will connect to them more cheaply. That is quiet news, but it is the kind of quiet news that shapes what your workday looks like a year from now.
❓ Frequently Asked Questions
What is MCP in simple terms?
MCP, the Model Context Protocol, is an open standard that lets AI assistants like Claude, ChatGPT, and Gemini connect to outside tools and data sources, such as your ad account, calendar, or online store. Anthropic introduced it in late 2024, and it has since become the common language for AI tool connections across major providers.
Does the MCP session removal break the AI tools I already use?
Not permanently. Providers are migrating their servers to the new 2026-07-28 spec, and the release candidate gave them a ten-week head start beginning May 21, 2026. You might see a connector hiccup during the transition, and reconnecting the integration usually fixes it. Once migrated, connections should be more reliable than before, not less.
Why does ad tech specifically have to rebuild for the new MCP spec?
Major ad platforms, including Google Ads, Amazon Ads, Meta, Snap, and Pinterest, built MCP servers around the old session model, which required sticky load balancer routing, shared session stores like Redis, and gateway-level request inspection. The stateless spec removes the need for all of that, so their infrastructure has to be reworked to match.
What are handles in the new MCP spec?
Handles are explicit identifiers, such as a basket_id or workflow_run_id, that a server returns directly to the AI agent instead of keeping state hidden in a session. The agent passes the handle back on later calls, which lets tasks survive dropped connections, be resumed later, or even be handed to a different agent.
🏁 Final Thoughts
The July 28, 2026 MCP update is a backend story with front-of-house consequences. Sessions are gone, handles are in, and every ad platform running an agent server is rebuilding to match. For solopreneurs and knowledge workers, the payoff is more reliable AI agents and a cheaper path for the tools you love to become agent-accessible. Keep one question handy for your vendors: are you on the 2026-07-28 spec yet? If this explainer saved you a research rabbit hole, subscribe to Agents at Work for weekly plain-English breakdowns of AI news, and drop a comment with the next headline you want decoded.
Last updated: August 01, 2026 · Keyword: MCP stateless update 2026 · Agents at Work
Comments
Post a Comment