Ditching "Sessions" and Going Back to Plain HTTP — MCP's Biggest Revision Yet Hands Long-Running Agents a Common Convention Called "Tasks"

MCP, the connectivity standard from Anthropic, reaches its biggest revision since launch on July 28, 2026. It removes "sessions" from the core to go stateless, and adds two extensions: "Tasks," a common convention for long-running work, and "MCP Apps," which return a UI. Breaking changes are include

Share
Ditching "Sessions" and Going Back to Plain HTTP — MCP's Biggest Revision Yet Hands Long-Running Agents a Common Convention Called "Tasks"

The Model Context Protocol (MCP), the connectivity standard that originated at Anthropic, has reached a milestone with the "2026-07-28" specification dated July 28, 2026. Its contents were locked down as a release candidate (RC) in May, and after roughly ten weeks of validation this date was set as the target for finalization. The official blog frames it as "the largest revision since launch." MCP is a prime example of a mechanism Claude introduced that has spread widely to other developer tools, and it has served as the foundation that connects coding agents to external tools and data. This revision reworks the very design philosophy of that foundation.

Toward "Stating Your Identity on Every Exchange" — Removing Sessions from the Core

The biggest change is the removal of "sessions" from the core of the protocol. Until now, a connection began with an initialize / initialized exchange (a handshake) and kept conversational state in a header called Mcp-Session-Id. The new specification abolishes both, and instead carries client information in a piece of accompanying data called _meta that is attached to every request.

The aim is operational ease. Because there is no state to hold (stateless), a request can be processed by whichever server instance receives it. Without "sticky sessions" that pin a conversation to a particular server, or any mechanism for sharing state, you can scale out horizontally behind an ordinary load balancer. The direction is to let remotely running MCP servers be operated with the same mindset as any ordinary web service.

A New Point of Entry Called Extensions — "MCP Apps" That Return a UI, "Tasks" That Make You Wait

The other pillar is the formalization of "Extensions." An extension carries a reverse-DNS-style identifier, negotiates its enabled/disabled state through capability exchange, and can be versioned separately from the specification itself. It is a framework for bolting on functionality while keeping the core thin, and two extensions ship together this time.

ExtensionWhat It Does
MCP AppsLets a server return an interactive HTML UI inside a sandboxed iframe. Tools declare their UI templates in advance, enabling prefetching and safety review. Interactions on the screen are handled over ordinary JSON-RPC
TasksA common convention for handling long-running work. In response to a tools/call, the server returns a "task handle," and the client drives its progress with tasks/get, tasks/update, and tasks/cancel

Tasks has been repositioned from what was previously an experimental core feature and redesigned as an extension. Instead of an immediate answer, the server returns a task ID, an initial state, and a recommended polling interval, and the client checks states such as working (in progress) / input_required (awaiting input) / completed / failed / cancelled via tasks/get. When input is needed partway through — as with waiting for human approval — values can be passed via tasks/update, allowing interactive, multi-step processing to proceed within a single task. It gives long-running agents a common way to "make you wait," so that each tool no longer implements it in its own scattered fashion.

Authentication Leans Toward OAuth, and Older Features Get a Notice of Departure

Authentication, too, has been brought closer to real-world operation. Six SEPs (specification proposals) align it with the practical use of OAuth 2.0 and OpenID Connect — for example, making validation of the iss parameter mandatory during authorization and requiring an application_type declaration for OpenID Connect. These are adjustments to make it easier to drop straight onto an enterprise identity platform.

Alongside this, the first formal deprecation policy has also been laid out. Roots, Sampling, and Logging are being deprecated, with a grace period of at least 12 months before removal. It is a set of rules for departure that lets things move forward without breaking.

The Two-Sided Nature: "It Works Even If You Don't Upgrade — But If You Do, Things Break"

For development teams, the benefits and the effort go hand in hand. This release includes breaking changes. Here are the main changes.

What ChangesDetails
SessionsAbolishes initialize / initialized and Mcp-Session-Id, going stateless
Error codesResource-not-found changes from -32002 to the standard -32602 (Invalid Params)
Tasks APIMigration from the old "2025-11-25" version becomes necessary
Legacy featuresRoots, Sampling, and Logging are deprecated (removal 12 months or more out)

Still, the official stance is that "existing servers keep working, and upgrading is optional and can proceed incrementally." To aid migration, beta SDKs for Python (v2.0.0b1), TypeScript (v2 beta), Go (v1.7.0-pre.1), and C# (v2.0.0-preview.1) have been released ahead of time, and Python and TypeScript come with migration guides that include automated conversion (codemods). That said, a caveat is attached: you should validate under real load behind a load balancer or gateway. This is a transitional period in which the standard is moving from a "local tool" to a "connectivity layer built for distribution," and the practical implication is that if you want to keep things running, early validation is required.

A standard that began with Claude is now becoming the plumbing shared by agents across many companies. This revision is not the flashy unveiling of a new model, but it is a move that shows how the conditions for "operating coding agents continuously and safely" are steadily solidifying outside of raw model performance.

References: The 2026-07-28 MCP Specification Release Candidate (Model Context Protocol Blog) / Beta SDKs for the 2026-07-28 MCP Spec Release Candidate / SEP-2663: Tasks Extension / MCP's Largest Spec Change Since Launch (TechTimes)

Read more

Making It Wait for "Jobs That Run Over an Hour": Codex 0.152 Adds Ceiling Dials for MCP Output Volume and Execution Time, and Turns the Planning Tool Off by Default

Making It Wait for "Jobs That Run Over an Hour": Codex 0.152 Adds Ceiling Dials for MCP Output Volume and Execution Time, and Turns the Planning Tool Off by Default

Codex v0.152.0 on August 31 and its next-day fix release added explicit ceilings on MCP tool output volume and execution time, and switched the planning tool off by default. Here's a rundown of the changes that matter for long-running unattended and semi-autonomous agent operation.

By FF
The CLI's Default Model Just Swapped In a Million-Token Brain — Claude Code v2.1.257 Makes Fable 5.1 the Standard and Adds a 'Containment Escape' Checkpoint to Auto Mode

The CLI's Default Model Just Swapped In a Million-Token Brain — Claude Code v2.1.257 Makes Fable 5.1 the Standard and Adds a 'Containment Escape' Checkpoint to Auto Mode

Claude Code v2.1.257, released September 1, 2026, swaps its default model to Fable 5.1 with its one-million-token context. It also adds guardrails to auto mode that stop credential retrieval and out-of-scope reads from slipping through. Here's a rundown of the changes that matter to developers.

By FF
"This Is an Authorized Exercise"—How the Aurora Ransomware Gang Insisted, While Making Cursor's AI Agent Do the Actual Intrusion Work

"This Is an Authorized Exercise"—How the Aurora Ransomware Gang Insisted, While Making Cursor's AI Agent Do the Actual Intrusion Work

Gambit Security and CloudSEK report that the ransomware group Aurora abused Cursor's AI agent for real intrusion work. Posing the tasks as an "authorized exercise" to slip past the safeguards, they had it handle reconnaissance and privilege takeover on the back of stolen credentials—a warning that a

By FF
One in Three Companies Now Choose to Build Rather Than Buy — McKinsey Measures How Coding Agents Are Reshaping the Procurement Decision

One in Three Companies Now Choose to Build Rather Than Buy — McKinsey Measures How Coding Agents Are Reshaping the Procurement Decision

McKinsey's annual survey found that about 30% of respondents passed on buying software because they could build it in-house with coding agents. We unpack the procurement shift from buying to building — and the current reality that productivity is up while profits stay flat.

By FF