AI Weekly: MCP Goes Stateless, Kimi K3, TSMC Records

Moonshot AI released Kimi K3, a 2.8 trillion parameter open-weight model with a 1 million token context window, on July 16. The model outperforms Western flagships on coding benchmarks and will have full open weights released on July 27 under a Modified MIT license. TSMC posted record earnings, confirming the accelerating hardware buildout for AI.

The week of July 11 to 18, 2026 delivered news at every layer of the AI stack. Moonshot AI shipped the largest open-weight model ever announced, Google targeted its long-delayed Gemini 3.5 Pro launch, and the Model Context Protocol published the biggest revision in its history. Underneath it all, TSMC posted record earnings that confirm the hardware buildout is still accelerating. Here is what happened, what the numbers say, and why it matters for people who build. A quick map of the issue. The coding tools section covers Kimi K3's agentic coding claims, the Gemini 3.5 Pro launch window, GitHub Copilot's July features, and the practical state of agent workflows. The processing section reads TSMC's record quarter, Intel's lithography first, and the low-precision formats reshaping inference cost. The standards section goes deep on the Model Context Protocol's stateless redesign, enterprise authorization, and the widening open-weight movement. Skim the headers if you only have five minutes. The details reward the full read. The coding tool story of the week started in Beijing. Moonshot AI released Kimi K3 on July 16 https://mlq.ai/news/moonshot-ai-releases-kimi-k3-a-28-trillion-parameter-open-weight-model-rivaling-top-us-systems/ , a 2.8 trillion parameter model built for long-horizon coding and agent workloads. The headline numbers are large. K3 carries a 1 million token context window, ships with reasoning always on, and includes native vision. Moonshot priced the API at $3 per million input tokens and $15 per million output tokens https://cryptobriefing.com/kimi-k3-open-weights-july-27/ , the highest pricing from any Chinese lab but roughly half the per-task cost of top Western frontier models. Full open weights land on July 27 under a Modified MIT license. Scale tells part of the story. K3 is roughly 2.8 times the size of its predecessor K2.6, and it dwarfs DeepSeek's 1.6 trillion parameter V4 Pro and Zhipu AI's 744 billion parameter GLM 5 series https://mlq.ai/news/moonshot-ai-releases-kimi-k3-a-28-trillion-parameter-open-weight-model-rivaling-top-us-systems/ . On the Artificial Analysis composite leaderboard, K3 posted an Elo of 1,547, a 732 point jump over the previous Kimi generation. Moonshot also reports that K3 uses 21 percent fewer output tokens than K2.6 on equivalent tasks https://mlq.ai/news/moonshot-ai-releases-kimi-k3-a-28-trillion-parameter-open-weight-model-rivaling-top-us-systems/ , which matters as much as raw capability when agents run thousands of steps per day. The company behind it has the funding to keep pushing. Moonshot, backed by Alibaba, Tencent, and Meituan, raised $2 billion at a $20 billion valuation in May and is reportedly in talks at a $30 billion valuation now. One legal cloud hangs over the launch: Anthropic accused Moonshot in February of training on 3.4 million Claude exchanges through distillation https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 , and K3 now benchmarks within a few points of the models named in that complaint. How that dispute resolves will shape the rules for every open-weight lab. The coding results explain the attention. Arena ranked K3 first in its Frontend Code evaluation at 1,679 points https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 in blind developer testing, ahead of every Western flagship. Moonshot's own evaluation suite places K3 behind Claude Fable 5 and GPT-5.6 Sol overall but ahead of everything else on coding and agentic benchmarks. One honest caveat belongs here. Every published K3 number is a Moonshot claim or drawn from API access https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 until the weights go public on July 27 and independent labs verify. Treat the rankings as promising, not proven. The verification gap cuts both ways, though. Once the weights publish, anyone can run the benchmarks, probe the failure modes, and fine-tune for their own stack. Closed models never face that level of scrutiny. K3 matters to coding tool users for a practical reason: Moonshot's models already power real developer products. Earlier Kimi versions were adopted by Cursor and DoorDash https://cryptobriefing.com/moonshot-kimi-k3-largest-open-weight-ai-model/ , so a stronger, cheaper Kimi flows straight into tools developers use daily. The race behind K3 has depth as well. Hong Kong listed MiniMax is building a 2.7 trillion parameter model of its own https://www.technology.org/2026/07/17/moonshot-kimi-k3-open-weight-ai-model/ , and Goldman Sachs began formally recommending Chinese models to Wall Street clients this year, a status shift that was unthinkable eighteen months ago. The 1 million token window fits whole repositories in a single prompt. The always-on reasoning mode has a cost, though. Independent testers measured 13,241 reasoning tokens for a simple SVG generation task https://mlq.ai/news/moonshot-ai-releases-kimi-k3-a-28-trillion-parameter-open-weight-model-rivaling-top-us-systems/ , about $0.25 for one query. Budget for thinking tokens if you route agent traffic to K3. Self-hosting math changes the calculus for large teams. Thanks to the quantization work covered in the processing section below, running K3 privately comes within reach of organizations holding 8 to 16 nodes of 8x H100 or B200 GPUs https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei . That is a serious cluster, and it is also a size that hundreds of enterprises and every national lab already own. A frontier-class coding model with no per-token bill and no data leaving the building is a new option on the menu, and the July 27 weights release is when the option becomes real. Google spent the week racing to its own launch. Google DeepMind targeted July 17 for Gemini 3.5 Pro general availability https://enterprisedna.co/resources/news/gemini-35-pro-july-17-rebuild-vs-deepseek-v4-2026/ after missing a June date. The delay had a dramatic cause. Google scrapped the original base model entirely and restarted pretraining after early testers flagged gaps in math, reasoning, and recursive tool calling. Circulating specifications describe a 2 million token context window, a Deep Think reasoning mode on the $250 Ultra tier, and pricing near $1.25 input and $10 output per million tokens. None of those specs came from official Google documentation https://www.techtimes.com/articles/320308/20260713/gemini-35-pro-targets-july-17-after-full-rebuild-every-spec-remains-unconfirmed.htm as of publication, so builders should wait for the model card before planning migrations. The rebuild story deserves a moment of respect. Shipping a flawed flagship on time is easy. Restarting pretraining six weeks before a promised date is expensive and embarrassing, and Google chose it anyway. While the Pro rebuild played out, Gemini 3.5 Flash carried production workloads since its May 19 launch https://www.techtimes.com/articles/320308/20260713/gemini-35-pro-targets-july-17-after-full-rebuild-every-spec-remains-unconfirmed.htm , posting 76.2 percent on Terminal-Bench 2.1 and 83.6 percent on MCP Atlas at $1.50 input and $9 output per million tokens. Flash handles the fast agent loops. Pro, when it lands, targets the hard reasoning at the top of the stack. Google has also been tuning the developer experience around its agent tooling, resetting and raising Gemini token quotas in its Antigravity coding product https://www.cnbc.com/2026/06/01/microsoft-and-google-take-on-anthropic-and-openai-in-ai-coding-models.html after developers burned through initial allocations faster than planned. Quota design sounds mundane, and it decides whether an agent product feels usable more than any benchmark does. If the 2 million token window is real, Gemini 3.5 Pro takes the context crown for whole-repo coding work. The launch calendar around it is the most crowded of the year. GPT-5.6 launched June 26 with its Sol, Terra, and Luna tiers, and Claude Fable 5 shipped June 9 https://memeburn.com/gemini-3-5-pro-targets-july-17-with-2m-token-context/ , so Gemini 3.5 Pro arrives last of the three frontier flagships. DeepSeek graduates its V4 family from preview to stable on July 24, the same week it retires legacy API aliases, which forces migration decisions on every team still pinned to old model names. On published coding benchmarks, Claude Fable 5 leads SWE-Bench Pro at 80.3 percent, against 58.6 percent for GPT-5.5 and 54.2 percent for the prior Gemini 3.1 Pro https://memeburn.com/gemini-3-5-pro-targets-july-17-with-2m-token-context/ . Gemini 3.5 Pro has no published score yet, and that empty cell in the comparison table is the one everyone wants filled. Google also renamed NotebookLM to Gemini Notebook https://www.buildfastwithai.com/blogs/ai-news-today-july-17-2026 this week, folding one of its most loved research tools into the Gemini brand as the whole product line consolidates. Microsoft shipped concrete updates rather than launch drama. The GitHub Copilot June update for Visual Studio https://github.blog/changelog/2026-07-14-github-copilot-in-visual-studio-june-update/ , published July 14, brings three features worth knowing. First, trust validation for MCP servers: Visual Studio now compares an MCP server's configuration and asset fingerprint against a trusted baseline at startup, and any change triggers a review dialog before the server runs. This lands right as MCP supply chain attacks became a serious research topic, and it is on by default. Second, the Copilot modernization agent for C++ graduated to general availability, handling MSVC upgrade scenarios end to end in automated mode or step by step in guided mode. Legacy C++ migration is exactly the kind of grinding, pattern-heavy work agents do well. The trust validation feature deserves a longer look because it models a discipline the whole ecosystem needs. An MCP server is executable capability handed to your model, and a poisoned update to a previously safe server is the classic supply chain move. Fingerprinting the approved configuration and interrupting on change is the same idea as lockfiles and signed packages, applied to agent tooling. Expect every serious MCP client to ship an equivalent within the year, and prefer the ones that already do. Third, long-distance next edit suggestions now predict follow-up edits anywhere in the active file, not just near the cursor, so a rename at the top of a file surfaces the matching fixes at the bottom. These features arrive against a changed business backdrop. Usage-based billing for GitHub Copilot went live for all users on June 1 https://github.com/orgs/community/discussions/192948 , with code review now consuming GitHub Actions minutes alongside AI credits. The flat-rate era of coding assistants is over across the industry, and every vendor is aligning price with token burn. The market they are fighting over keeps growing. Mordor Intelligence projects AI code tools expanding 26 percent a year, from $9.3 billion in 2026 to roughly $30 billion by 2031 https://www.cnbc.com/2026/06/01/microsoft-and-google-take-on-anthropic-and-openai-in-ai-coding-models.html . Developer sentiment data shows where loyalty sits right now. The Pragmatic Engineer survey from February named Claude Code the most loved tool at 46 percent, against 19 percent for Cursor and 9 percent for GitHub Copilot, and found 70 percent of teams running two to four AI tools in parallel https://pasqualepillitteri.it/en/news/3392/github-copilot-cursor-claude-code-ai-coding-showdown-2026 . Nobody standardized on one assistant. Teams compose stacks, with terminal agents for deep tasks, IDE assistants for daily edits, and cloud agents for background work. What does a working developer do with all of this? A few practical takeaways from the week. First, revisit your token budgets. With usage-based billing spreading and always-on reasoning models burning thousands of thinking tokens per request, the cost profile of your agent workflows changed this quarter even if your code did not. Measure cost per completed task, not cost per token, and route easy work to cheap fast models. Second, treat MCP server trust as a real attack surface. Visual Studio's fingerprint validation is a template worth copying anywhere you run third-party tool servers: pin what you approved, and alert on drift. Third, hold one slot in your evaluation harness for open-weight challengers. If K3's numbers survive independent testing after July 27, self-hosted frontier coding assistance becomes a line item you can price against API bills, and procurement conversations change fast when that line item exists. Agent infrastructure kept maturing around the editors. Perplexity introduced Secure Sandboxes https://johnsviokla.substack.com/p/ep-622-daily-ai-news-july-17-2026 on July 17, an isolation layer that gives autonomous agents contained execution environments, credential management, and hard security boundaries. The launch answers the question every platform team asks before approving agent deployments: what happens when the agent runs code we did not review? Replit engineers reported tripling code output https://radicaldatascience.wordpress.com/2026/07/17/ai-news-briefs-bulletin-board-for-july-2026/ using an internal system of coordinated AI agents, a data point for the multi-agent workflow pattern that spread through 2026. The pattern has a recognizable shape now. Cursor, Claude Code, and OpenAI Codex stopped converging into one winner and instead formed layers of a composable stack https://thenewstack.io/ai-coding-tool-stack/ : one tool orchestrates parallel agents, another executes deep changes, and a third reviews asynchronously in a cloud sandbox. The review layer follows a sound principle: asking the model that wrote code to also review it means grading its own homework, so teams route review to a different system on purpose. Sandboxing products like Perplexity's slot straight into that stack as the execution containment layer, giving each agent an isolated environment with scoped credentials so a misbehaving step damages nothing outside its box. And at Google Cloud Next in Las Vegas, Sundar Pichai said close to 75 percent of code at Google is now AI generated and engineer approved https://local.newsbreak.com/trending/top/ai-in-coding-news , up from 25 percent in 2024 and 50 percent in 2025. He described engineers orchestrating autonomous agent fleets, and cited a code migration that agents finished six times faster than human teams. Numbers like that from a company of Google's size move the baseline for everyone. When three quarters of a major engineering organization's code arrives machine-written, the scarce skills shift to specification, review judgment, and system design, and hiring plans across the industry are already adjusting to match. Put the week together and a pattern appears. The frontier labs compete on reasoning depth and context length. The tool vendors compete on trust, isolation, and workflow fit. Both layers moved this week, and the gap between a raw model and a production coding agent keeps widening. That gap is where the interesting engineering lives. One more data point tempers the enthusiasm. GitLab's 2026 AI Accountability Report found 78 percent of developers report faster code output and 73 percent report better quality, yet overall software delivery has not sped up https://www.infoq.com/news/2026/06/ai-coding-outpaces-governance/ , because testing, review, and governance bottlenecks absorb the gains downstream. The report frames the fix as accountability: for any line of AI-generated code, an organization should be able to answer where it came from, what it was meant to do, and who approved it. Only a third of surveyed organizations that suffered an incident in the past year were able to trace whether AI-generated code contributed. Writing code faster was never the whole job. Shipping trustworthy systems is, and that is where 2026's tooling investments are heading. If you want one number that summarizes the state of AI hardware demand, TSMC provided it on July 16. The world's largest contract chipmaker reported a 77 percent surge in net income and record quarterly revenue https://www.distillintelligence.com/briefings/semiconductors-ai-chips-2026-07-17 , driven by AI chip orders from Nvidia, AMD, Apple, and the hyperscalers. The run-up told the same story. June sales alone jumped 68 percent year over year https://www.fxleaders.com/news/2026/07/13/tsmc-hits-record-highs-as-ai-chip-monopoly-powers-relentless-rally/ , quarterly sales rose 36 percent, and the company's 3 nanometer and 5 nanometer capacity is fully booked through 2027. Advanced packaging, the CoWoS step that joins compute dies to high-bandwidth memory, remains the binding constraint on AI chip supply, and TSMC keeps expanding it against a backlog. Three new advanced packaging facilities in Chiayi project more than 300 billion Taiwan dollars in annual output once ramped. The market treated the report as a referendum on the whole AI trade. TSMC stock is up more than 52 percent in 2026, and analysts framed the quarter as a health check for trillions of dollars of AI-linked market value. The demand signal from TSMC's largest customer backs it up. Nvidia's most recent quarter delivered $81.61 billion in revenue, up 85.2 percent year over year https://247wallst.com/investing/2026/07/13/tsmc-sales-jump-36-as-memory-stocks-plunge-what-it-means-for-nvidia-and-amd/ , with demand spread across AI labs, hyperscalers, sovereign programs, and the new tier of GPU cloud providers. Not every corner of the hardware market shared the party. Memory stocks slumped in the same week, with SK Hynix falling 13 percent in one session https://247wallst.com/investing/2026/07/13/tsmc-sales-jump-36-as-memory-stocks-plunge-what-it-means-for-nvidia-and-amd/ on oversupply worries. The split matters for anyone reading the boom: logic capacity for AI accelerators remains supply-constrained while parts of the memory market wobble, so "AI hardware" is no longer one trade. Follow the pricing thread to its end and the industry structure comes into focus. TSMC raises wafer prices because it can, since rivals trail on yield at the leading edge. Chip designers pass the increase to cloud providers, who pass it to AI companies, who face a choice: raise API prices, eat margin, or engineer the cost out. That third option explains half of this newsletter. Custom silicon programs, sparse architectures, low-precision formats, and token-thrifty models are all the same answer to the same invoice. Compute scarcity became the field's chief designer, and the designs are getting good. TSMC paired the earnings with a capital announcement that reshapes the map. The company will add $100 billion to its Arizona manufacturing investment, bringing the total there to $265 billion https://www.fool.com/investing/2026/07/16/tsmc-just-announced-fantastic-news-for-nvidia-shar/ . For US chip customers, that number converts geopolitical risk into concrete fab capacity on American soil over the coming years. TSMC also told major customers to expect wafer price increases of 5 to 10 percent https://www.thestreet.com/investing/stocks/chip-trade-waiting-on-taiwan-semiconductor-tsm-report , and the hikes now reach beyond the newest 3 nanometer node. Pricing power flows downstream. Expect it in GPU prices, then in cloud instance rates, then in your inference bill. The equipment layer confirmed the boom. ASML raised its full-year 2026 sales forecast https://www.distillintelligence.com/briefings/semiconductors-ai-chips-2026-07-17 after quarterly earnings beat expectations on a surge of orders for its lithography machines. The more striking ASML story came from its biggest new customer. Intel became the first company to ship high-volume logic chips built with ASML's High-NA EUV scanners https://www.distillintelligence.com/briefings/semiconductors-ai-chips-2026-07-17 , with select Panther Lake layers on the 18A node now qualified for the 0.55 NA machines. Reports also point to major yield gains on 18A, with figures around 85 percent circulating. High-NA EUV prints finer features in fewer steps, and every leading-edge roadmap depends on it. Intel reaching volume production first, after years of trailing TSMC on process, is the strongest signal yet that its foundry comeback has substance. Dual qualification is the detail worth understanding: the same Panther Lake layers now print on both the standard 0.33 NA machines and the new 0.55 NA scanners, so Intel can shift volume between tool fleets and prove the new machines against a known baseline. Reports of Intel Foundry winning fresh chip orders followed the announcement within days. For AI buyers, a second credible leading-edge foundry means pricing pressure on the incumbent and resilience against a single point of geographic failure, both outcomes the industry has wanted for a decade. ASML is now preparing TSMC and Samsung for their own High-NA waves. A quick decoder for readers outside the fab world: EUV lithography uses extreme ultraviolet light to print chip features, and the numerical aperture of the optics sets how fine those features get. The new 0.55 NA machines, at roughly $400 million each, print smaller transistors in a single exposure where older tools need several. Whoever masters them first gets density and cost advantages that compound for years, which is why Intel's milestone reached far beyond one product line. Nvidia spent the week expanding sideways. The company launched Cosmos 3 Edge, a vision reasoning model for edge deployment, and deepened its partnership with Japan https://www.distillintelligence.com/briefings/semiconductors-ai-chips-2026-07-17 to build national AI infrastructure on next-generation Rubin chips. Sovereign AI, governments buying their own training capacity, has become a durable demand pillar alongside the hyperscalers. Japan's program pairs national compute with domestic robotics and manufacturing data, and Nvidia's edge push fits the same thesis. Cosmos 3 Edge targets vision reasoning on devices in factories, vehicles, and stores, where round trips to a distant data center cost too much latency. Training stays centralized. Inference is spreading to wherever the cameras are, and the chip demand curve now has two humps, one in the data center and a growing one at the edge. Policy moved in parallel. The United States approved shipment of a limited number of advanced AI chips to select Chinese buyers https://www.distillintelligence.com/briefings/semiconductors-ai-chips-2026-07-17 , even as Nvidia reportedly cut its Asian buyer list in half to tighten export compliance. The export regime is turning from a wall into a valve, opened and closed buyer by buyer. The workaround economy on the other side keeps growing too. A Huawei-led team reported post-training DeepSeek's 1.6 trillion parameter model on 1,000 Ascend 910C chips https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 , proof that frontier-scale work now happens on domestic Chinese silicon when imports fall short. The model layer answered the hardware layer with an argument about compute itself. Kimi K3's engineering choices read like a manifesto for doing more with less. The model activates just 16 of its 896 experts per token, about 1.8 percent of its parameter pool https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 , so a 2.8 trillion parameter model runs at a fraction of dense-model cost. Its Kimi Delta Attention design decodes up to 6.3 times faster than standard attention https://cryptobriefing.com/kimi-k3-open-weights-july-27/ , and an Attention Residuals technique improved training throughput about 25 percent over the prior generation. Moonshot also started quantization-aware training at the supervised fine-tuning stage, using MXFP4 weights and MXFP8 activations https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 . Those are open microscaling formats, and baking them in during training rather than compressing afterward is how a 2.8 trillion parameter model becomes self-hostable on clusters of 8 to 16 nodes of H100 or B200 GPUs https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei . K3 even posted results on GPU kernel generation, sustaining more than 8,700 tokens per second of simulated decode in one chip-design benchmark. A word on those formats, since they are becoming vocabulary every data engineer needs. MXFP4 and MXFP8 are microscaling number formats standardized through the Open Compute Project, storing blocks of values at 4-bit or 8-bit precision with a shared scaling factor per block. They cut memory footprint and bandwidth needs by half or more compared to 16-bit weights, and modern accelerators execute them natively. Training with the target precision from the start, instead of quantizing a finished model, preserves quality that post-hoc compression loses. Chinese labs, squeezed by export controls, are turning compute scarcity into architecture research, and the whole field inherits the results when the weights open. One more silicon thread continued from the start of the month. Anthropic remains in early talks with Samsung https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/ about manufacturing a custom AI accelerator, first reported July 2, with Samsung's 2 nanometer process under evaluation. OpenAI already unveiled its Broadcom-built inference chip, Jalapeno. Every frontier lab now treats custom silicon as a lever on inference cost, and the foundry earnings above show why: the bill for rented compute keeps climbing, and 10 to 30 percent inference savings from purpose-built chips changes the economics of serving models at scale. Anthropic's broader infrastructure commitments give the talks context: the company has committed more than $100 billion in AWS purchases and a $50 billion US data center buildout with Fluidstack, and it recently hired Clive Chan, an engineer from OpenAI's silicon program, a signal the chip project has moved past idle exploration. The week's processing news fits one frame. Demand is verified and rising, per TSMC and ASML. Supply is diversifying, per Intel's High-NA milestone and Arizona's buildout. And the software side is attacking the same problem from above, with sparse activation and low-precision formats cutting the compute each token needs. Cost per unit of intelligence is the metric every one of these stories moves. The Model Context Protocol, the open standard that connects AI models to tools and data, published the release candidate for its 2026-07-28 specification https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ this week. The maintainers call it the largest revision since the protocol launched in November 2024, and the changes read like a graduation from promising project to production infrastructure. The trajectory to this point was fast even by AI standards. Anthropic introduced MCP twenty months ago as a universal way for models to reach tools and data. OpenAI and Google DeepMind adopted it within months, an almost unheard-of alignment among direct competitors, and the server ecosystem grew from dozens to thousands. Growth exposed the seams: session state that fought load balancers, authorization that predated enterprise identity practice, and a core spec absorbing every new idea. The 2026-07-28 revision addresses all three at once. The centerpiece is a stateless core. The revision removes the initialize handshake and the protocol-level session entirely https://4sysops.com/archives/2026-07-28-model-context-protocol-mcp-stateless-multi-round-trip-routable-headers-authorization-hardening/ . Every request now travels self-contained, carrying the protocol version, client information, and capabilities instead of relying on state exchanged up front. For anyone who has operated a remote MCP server, this solves the deployment headache directly. A server that previously needed sticky sessions and a shared session store can now run behind a plain round-robin load balancer https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ . New Mcp-Method and Mcp-Name headers let gateways route traffic without inspecting request bodies, which unlocks clean rate limiting and service mesh integration. New ttlMs and cacheScope fields on list and read responses give clients defined caching rules, so a tool list gets fetched once and reused safely instead of hammered on every turn. The revision also introduces multi-round-trip request patterns, so a single logical operation can span several exchanges without resurrecting session state. That combination, stateless transport plus structured multi-step interactions, is what lets MCP serve both a laptop-local tool server and a fleet of containers behind a global load balancer with the same specification. Statelessness is the boring-sounding change that decides whether a protocol survives contact with production traffic. HTTP won the web partly because any server can answer any request. MCP just adopted the same survival trait. The revision also restructures how MCP grows. Capabilities like server-rendered user interfaces, called MCP Apps, and long-running work, called the Tasks extension, now live as first-class extensions that ship on their own timelines https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ rather than bloating the core. Authorization moved closer to the OAuth and OpenID Connect deployments enterprises already run. A formal deprecation policy commits the project to evolving without breaking existing builds. Tool definitions gain full JSON Schema support, so complex parameter shapes finally validate the same way everywhere. The two flagship extensions deserve their own sentences. MCP Apps lets a server return rendered user interface components, so a tool can hand back an interactive chart or form instead of raw JSON for the client to guess at. The Tasks extension standardizes long-running work: an agent kicks off a job, polls or subscribes for progress, and collects results later, the pattern behind research runs, batch data jobs, and slow external APIs. Pulling these out of the core means a minimal server stays minimal while ambitious servers grow capabilities on a published track. The formal deprecation policy is the quiet companion to all of it. Enterprises refused to build on a protocol that changed under their feet, and a written lifecycle for retiring features is the price of their trust. The tooling is ready to test today. Beta releases of the Python, TypeScript, Go, and C SDKs shipped alongside the release candidate https://4sysops.com/archives/2026-07-28-model-context-protocol-mcp-stateless-multi-round-trip-routable-headers-authorization-hardening/ . Python v2 renames FastMCP to MCPServer while keeping the decorator API. TypeScript v2 splits the single package into focused modules for server and client, and goes ESM-only. Go ships support in v1.7.0-pre.1 and C in a 2.0.0 preview. Compatibility is handled gracefully: new clients fall back to the old handshake when they meet a server on an earlier revision, so nothing breaks on July 28 when the final specification publishes. If you maintain an MCP server, start validating against the release candidate now, and if you use the Tasks pattern, plan the migration to the extension-based lifecycle. The final specification publishes July 28, and Tier 1 SDKs are expected to ship full support within a ten-week validation window https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ under the project's SDK tier system. The changelog lists every change against the 2025-11-25 revision, and the specification repository takes issues from implementers who hit problems. The stateless release lands on top of an enterprise security milestone from earlier this month that deserves mention for anyone rolling out MCP at work. The project promoted its Enterprise-Managed Authorization extension to stable status https://www.infoq.com/news/2026/07/mcp-ema-enterprise-auth/ , replacing per-server consent prompts with a single sign-on flow through the organization's identity provider. Users authenticate once, and approved servers just work. The flow rides the identity provider an organization already operates, so access reviews, offboarding, and audit trails cover MCP servers the same way they cover every other SaaS application. That single property converts MCP from a tool individual developers sneak past IT into a system IT can approve. Anthropic, Microsoft, and Okta adopted the extension, and it gives IT departments the central control they require before approving agent tooling. Pair that with Visual Studio's new MCP trust validation, covered above, and a theme emerges: the ecosystem spent this cycle hardening the protocol for organizations, not just enthusiasts. Adoption breadth keeps compounding. Microsoft's MCP catalog now includes more than 60 ready servers https://www.microsoft.com/en-us/power-platform/blog/2026/07/06/dataverse-july2026/ spanning its productivity, developer, and business application stack, usable across Microsoft 365 Copilot, Copilot Studio, Azure AI Foundry, and GitHub Copilot. One standard connection model across all of those surfaces is exactly the outcome protocol standardization promised. Consumer platforms keep joining too. X launched a hosted MCP server https://www.techbuzz.ai/articles/x-launches-mcp-server-to-bridge-ai-apps-and-platform-api that opens its platform API to AI applications through the standard interface, sparing developers custom integration work. When social platforms, enterprise suites, and developer tools all speak one protocol, agent builders stop writing adapters and start writing behavior. The enterprise agent platforms racing to consume these standards showed their strategy this week as well. Google is selling enterprises the tooling to deploy fleets of governed agents https://www.buildfastwithai.com/blogs/ai-news-today-july-14-2026 that connect to corporate data, run multi-step workflows, and stay under IT control, and both Google and Microsoft are backing shared standards for how agents connect to business software. The operative word in every enterprise pitch is govern. Companies stall on agents not because models are weak but because unmanaged agents leak data and exceed authority. Standards plus governance is the unlock, and this week delivered progress on both halves. Agent-to-agent communication had its own moment of maturity, in the form of clear-eyed security writing. The Agent2Agent protocol, stewarded by the Linux Foundation, passed 150 supporting organizations with production deployments across multiple industries https://www.glukhov.org/ai-systems/comparisons/a2a-protocol-2026-adoption/ as of April. A detailed security analysis published July 16 https://arnav.au/2026/07/16/securing-agent-to-agent-a2a-communication/ walked through what A2A deliberately leaves to deployers: identity, credential provisioning, and authorization sit outside the protocol, and closing that gap is the operator's job. The protocol runs on JSON-RPC 2.0 over HTTPS with Server-Sent Events for streaming, and Agent Cards advertise capabilities for discovery. The division of labor across the standards is now settled shorthand: MCP connects agents to tools, A2A connects agents to each other, and security teams own the identity layer both standards ride on. The concrete risks the analysis names are worth internalizing before your first multi-agent deployment. An agent that trusts another agent's self-description trusts an unverified claim, so capability discovery needs authentication behind it. Delegated tasks carry data across trust boundaries, so payloads need classification and filtering, not just encryption in transit. And long-running agent relationships need credential rotation and revocation, because a compromised agent with standing delegations is a lateral movement machine. None of this is a flaw in A2A. It is the ordinary work of operating any federated system, arriving in a new costume. Licensing counts as a standard too, and the week produced a notable data point. Moonshot's decision to release Kimi K3 under a Modified MIT license https://cryptobriefing.com/kimi-k3-open-weights-july-27/ keeps the largest open-weight model ever announced permissive enough for commercial use, fine-tuning, and integration without legal friction. Open weights at frontier scale change who gets to build serious AI systems. Enterprises with their own GPU clusters, researchers probing model internals, and startups fine-tuning for narrow domains all gain an option that no closed API offers. July 27, when the weights actually drop, will test whether the community can reproduce the benchmark claims. Watch that date. Open weights and open protocols reinforce each other, which is why they share a section. A team that self-hosts K3 still needs its agents to reach tools, and MCP is how they will do it without vendor lock-in at the integration layer. A company standardizing on MCP gains the freedom to swap models, closed or open, without rewriting a single connector. Each open layer raises the value of the others, and the stack that results, open model weights over open protocols over open data formats, is the same architectural bet the lakehouse world made about data a decade ago. Portability wins slowly, then all at once. The calendar for the coming ten days is unusually dense. The World Artificial Intelligence Conference in Shanghai runs through July 20 with more than 140 forums, after opening with Xi Jinping's first keynote in the event's history, and Chinese labs traditionally time releases to it. July 24 brings DeepSeek's V4 stable graduation and the retirement of its legacy API aliases, a forced migration for anyone still on old model names. July 27 is the K3 open weights drop, when independent benchmarking begins in earnest. And July 28 is the MCP specification final, the starting gun for the SDK support window. Any one of these reshapes a corner of the stack. All four in one stretch make the last week of July a checkpoint for the whole year. Set your evaluation pipelines up before the dates hit, not after. Teams that had harnesses ready when GPT-5.6 and Fable 5 launched in June made routing decisions in days. Teams that started building on launch day are still catching up, and the gap between those two groups is becoming a real competitive difference in how fast organizations absorb new capability. The releases will keep coming. The absorption machinery is the durable investment. Every section above tells the same story from a different altitude. The protocol layer went stateless so agent infrastructure scales like ordinary web infrastructure. The model layer used sparsity and low-precision formats to cut the compute behind each token. The hardware layer posted record numbers while adding capacity on two continents. The industry is industrializing. The experiments of 2024 and 2025 are becoming load-bearing systems with load-bearing standards, and the winners of the next phase will be the teams that treat agents, models, and data as one engineered stack rather than three separate bets. For data teams specifically, the assignment is clear. Agents are becoming the primary consumers of analytical data, standards now define how they connect, and the economics reward architectures that keep data open, governed, and queryable by any model you choose next year. Build for that world now and the next model launch becomes a config change instead of a migration. The AI world changes fast. Here are tools and resources to help you keep pace. Try Dremio Free : Experience agentic analytics and an Apache Iceberg-powered lakehouse. Start your free trial https://www.dremio.com/get-started?utm source=ev external blog&utm medium=influencer&utm campaign=pag&utm term=07-18-2026&utm content=alexmerced Learn Agentic AI with Data : Dremio's agentic analytics features let your AI agents query and act on live data. Explore Dremio Agentic AI https://www.dremio.com/use-cases/agentic-ai/?utm source=ev external blog&utm medium=influencer&utm campaign=pag&utm term=07-18-2026&utm content=alexmerced Join the Community : Connect with data engineers and AI practitioners building on open standards. Join the Dremio Developer Community https://developer.dremio.com/?utm source=ev external blog&utm medium=influencer&utm campaign=pag&utm term=07-18-2026&utm content=alexmerced Book: The 2026 Guide to AI-Assisted Development : Covers prompt engineering, agent workflows, MCP, evaluation, security, and career paths. Get it on Amazon https://www.amazon.com/2026-Guide-AI-Assisted-Development-Engineering-ebook/dp/B0GQW7CTML/ Book: Using AI Agents for Data Engineering and Data Analysis : A practical guide to Claude Code, Google Antigravity, OpenAI Codex, and more. Get it on Amazon https://www.amazon.com/Using-Agents-Data-Engineering-Analysis-ebook/dp/B0GR6PYJT9/ Browse the full catalog of 50+ books at books.alexmerced.com https://books.alexmerced.com