nextbig.dev
Vancouver, B.C. · Intelligence on AI and the machines that run it
nextbig.dev
The Briefing · Tuesday, July 28, 2026

MCP removed its handshake and session state after reaching nearly half a billion monthly SDK downloads, turning agent tools into ordinary routable web infrastructure

The 2026-07-28 specification makes every request self-describing, routable through plain load balancers and cacheable at the catalog layer. TypeScript and Python have each crossed 1bn cumulative SDK downloads, so a breaking migration away from session IDs now changes production infrastructure rather than an experimental interface.

7 min read
The Rundown No. 159 · Audio Edition · 2 min All episodesRSSMP3
0:00 / 1:48
VTT
The Big Story
MCP removed its handshake and session state after reaching nearly half a billion monthly SDK downloads, turning agent tools into ordinary routable web infrastructure

Close to half a billion Tier 1 SDK downloads now happen in a month, and the protocol behind them just deleted its session. MCP's 2026-07-28 specification retires the required initialize exchange and the Mcp-Session-Id header. Each request carries its protocol version, client identity and capabilities, so it can land on any instance behind a plain round-robin load balancer without shared transport state. TypeScript and Python have each crossed 1bn cumulative downloads. A design decision inside an 18-month-old protocol now reaches an installed base measured like a mainstream web library.

That sounds like plumbing because it is. The earlier stateful design made a remote tool server behave unlike the web infrastructure teams already know how to operate. Sticky sessions, held-open streams and hidden connection state complicate scaling, failover and debugging. The new core moves durable application state into explicit handles that a model can pass between tools. A crashed server becomes another request to another instance, while the state the agent needs stays visible in the conversation and logs. Ordinary autoscaling and blue-green deployment become available without inventing a parallel control plane for agents.

Routing and caching move up a level with it. Method and tool names now travel in Mcp-Method and Mcp-Name headers, letting a gateway, rate limiter or firewall authorize and meter a call without parsing the JSON body. Tool, prompt and resource lists gain deterministic order plus ttlMs and cacheScope hints. Multi Round-Trip Requests handle approvals or missing input by returning an input_required result and retrying the original call, replacing several server-initiated requests that needed an open stream. The approval remains attached to the request that caused it.

The migration is genuinely breaking for teams that stored meaning in session identifiers. Roots, Sampling and Logging are deprecated, as is the legacy HTTP plus SSE transport, although the project promises at least a 12-month offramp. Dynamic Client Registration is also headed out in favor of client metadata documents. The security work includes RFC 9207 issuer validation and binding credentials to the authorization server that minted them, closing room for an authorization-server mix-up. Implementers get a year, but new servers should stop accumulating session-dependent behavior now under real production load.

The four Tier 1 SDKs for TypeScript, Python, Go and C# already support the revision. AWS says the stateless core is available in Bedrock AgentCore, Cloudflare supports it from day zero, and Microsoft Foundry is using MCP as the route from dozens of integrations to thousands. That distribution makes this more than a neat protocol correction. The useful test is operational: remove affinity from an MCP service, send successive tool calls to different instances, and verify that approval, authorization and explicit state still survive. If they do, agent infrastructure finally fits the same load balancer runbook as every other HTTP workload.

@modelcontext Read source
Security Starts Routing by Job

Microsoft and Wiz route security work across models and clear 90% on CyberGym

Microsoft and Wiz reported that the Atlas multi-model system found 90.9% of vulnerabilities on CyberGym and more than 200 previously unknown flaws, while Microsoft's MDASH reached 95.95% on its test. Individual models were closer to 83% to 86%. Atlas sends roughly 90% of work through a smaller model and escalates the difficult remainder, which the companies say halves cost. That is a useful mechanism: cheap triage, specialized review and proof before a result ships. Atlas remains an internal system rather than a product, however, and benchmark recall omits the operational price of false positives, duplicate findings and fixes that cannot be reproduced.

Spur raises $200m as bots become the majority of internet traffic

Bot-detection company Spur raised $200m from Insight Partners after automated traffic reportedly exceeded human traffic in mid-2026. Founded in 2017, the company classifies the networks and infrastructure behind requests rather than treating every unusual browser event as equivalent. Agent traffic makes that distinction harder and more valuable because a legitimate purchasing agent, a scraper using a residential proxy and credential-stuffing malware can all look automated while requiring different policy. MCP's named method headers help at the application layer, but public endpoints still need network reputation, identity and intent signals. The funding round prices a shift from blocking automation to deciding which machines are allowed to act.

The Spending Meets Evidence

Google raises its capital plan again while investors ask for the utilization line

Alphabet raised its 2026 capital-spending guide to $195bn to $205bn from $180bn to $190bn, adding another $15bn at both ends after an already aggressive build plan. The spending can be rational while contracted demand and cloud growth stay strong, but the disclosure still leaves the key operating number out: how much newly energized capacity is productively used. Revenue is a lagging proxy because reservations, construction and deployment arrive on different clocks. Builders should watch cloud margin, free cash flow and leased-capacity reliance together. A higher guide with stable margins says demand is absorbing the infrastructure; another increase alongside weaker cash generation says the queue is driving the plan.

Google's 14.65m-interaction study finds AI work is still collaboration more than automation

Google analyzed 14.65 million deidentified workplace AI interactions for its ATLAS research and found that most use remains collaborative, with full automation uncommon. That result is a useful counterweight to capital plans built around agents replacing whole workflows. Assistance can still justify substantial compute if it reaches enough workers, but the unit of value is often a draft, explanation or decision supported rather than a job completed. It also changes measurement. Seat activation and message volume say little about whether the system closes work. Teams should log completed outcomes, human corrections and elapsed time at the task boundary, especially as MCP makes tool calls easier to scale.

Quick Hits
The Takeaway

MCP's revision and Microsoft's security results both move intelligence out of a long-lived conversation and into explicit routing. The protocol now exposes method names to gateways, carries state in tool-visible handles and lets any instance serve the next request. Atlas sends most vulnerability work to a smaller model and escalates the remainder. Spur's $200m round prices the public-network version of the same problem: machines are now the majority of traffic, so infrastructure has to identify the job and authority behind each request. The operational gain is ordinary, which is why it matters. Stateless services can use standard load balancers, named calls can be metered, and task outcomes can be measured independently of a model session. Google's 14.65m-interaction study says most workplace use is still collaborative, so teams should resist equating more protocol traffic with more completed work. Migrate one MCP service without affinity, route by method, attach an outcome and cost to each call chain, and see whether the production system survives the abstraction.

The Call C-20260728

By October 31, 2026, at least one major commercial code-security vendor will generally release an auditable production scanner that routes discovery and verification across multiple models and publishes a result above 90% on a named public vulnerability benchmark.

The case

Microsoft and Wiz have shown a routed system crossing 90% while individual models remain several points lower, and the cheaper model handles about 90% of Atlas traffic. That creates a quality gain and a cost story a security vendor can sell together. MCP's header-level routing and stateless calls reduce the infrastructure novelty. The remaining work is productization: reproducible evidence, false-positive controls and a public result customers can inspect.

What proves us wrong

If October 31 arrives with Atlas still internal and no major vendor generally offering a multi-model routed scanner with a published score above 90% on a named public benchmark, the call is wrong.

Settles by October 31, 2026
The Tape T-20260728
◆ Watch MSFT Microsoft medium conviction

We sharpen the Microsoft watch for a second day. Atlas and MDASH turn yesterday's model-routing argument into measured security work: more than 90% vulnerability recall, over 200 newly found flaws and a design that sends most traffic to the cheaper engine. MCP's stateless release also strengthens Foundry's integration layer. The missing piece is a sellable product with public false-positive and cost data.

Microsoft can combine model routing, identity and a widely adopted tool protocol inside Foundry. The internal scores are strong, but commercial value depends on packaging the system and proving that benchmark gains survive production repositories.

Wrong if Atlas remaining internal through October, or a public evaluation showing routed gains disappear after false positives and remediation failures are counted, would weaken the product case. Settles 6 months
◆ Watch GOOGL Alphabet medium conviction

We hold the Alphabet watch. The new $195bn to $205bn capital guide expands the cash-flow concern from July 24, while the ATLAS study supplies a sober demand measure: 14.65m interactions still skew toward collaboration rather than complete automation. Assistance at Google scale can fill infrastructure, but the company must show productive utilization before another guide increase turns contracted demand into a perpetual construction explanation.

The capex range rose by $15bn at both ends, while Google's own evidence says workplace AI remains mostly collaborative. Cloud growth and backlog can justify the build if margins and cash conversion hold.

Wrong if Stable or expanding cloud margins with free cash flow returning positive despite the higher guide would retire the utilization concern; another increase alongside weaker cash generation would confirm it. Settles 9 months
Desk signals from the day's verified wire — falsifiable, dated, settled in public. Analysis, not individualized investment advice.

Get this briefing in your inbox

What changed in AI and compute, what it costs, and what to build. One email per week. No spam, unsubscribe anytime.