# AI Weekly: Qwen4 Preview, Hot Chips, and Agent Tools Go GA

> Source: <https://dev.to/alexmercedcoder/ai-weekly-qwen4-preview-hot-chips-and-agent-tools-go-ga-3cn7>
> Published: 2026-08-27 15:21:34+00:00

*Week of August 19 to 26, 2026*

The labs took a breath on flagship releases this week, and the hardware people filled the silence. Alibaba shipped an open-weight preview of its next architecture. Anthropic moved its agent tooling out of beta. Google's agent protocol changed foundations. And at Hot Chips, Nvidia, Google, and OpenAI each showed a chip designed around one idea: agents generate a lot of tokens, and the decode phase is where the money goes.

The most consequential model release of the week is not a flagship. Alibaba's Qwen team [released Qwen3.8-Flash-Next on August 26](https://github.com/QwenLM/Qwen3.8-Flash-Next/), an open-weight multimodal mixture-of-experts model that the team describes as an early preview of the architecture Qwen4 will be built on. The team drew a direct parallel to Qwen3-Next, which introduced the Gated DeltaNet plus Gated Attention design that then carried through the Qwen3.5, 3.6, 3.7, and 3.8 series. Flash-Next plays the same role for Qwen4: release the architecture early so the community can study it before the full model family arrives.

The numbers describe an unusual shape. The main model carries 125 billion parameters, but only 6 billion are active per token. On top of that sits a separate 51 billion parameter N-gram embedding layer. That layer stores common word groups as standalone entries in what The Decoder called a phrase dictionary, and it can sit in ordinary system RAM rather than on the accelerator, which is a way to add capacity without paying for it in GPU memory or compute. The model supports a native context window of 262,144 tokens and extends to roughly one million with YaRN. Alibaba says training cost about one-ninth of what Qwen3.7-Plus cost, and Qwen3.7-Plus is a 397 billion parameter model with 17 billion active.

The architecture changes span four areas. Attention pairs Gated DeltaNet with a new Qwen Sparse Attention that selects context at the level of micro-blocks rather than individual tokens, aimed at cutting latency on the long prompts that dominate agent workloads. The residual stream adds separate read and write gates. The embedding layer is the N-gram addition. And the training recipe splits the Muon and AdamW optimizers across different weight categories and starts training at the target batch size instead of warming up to it. Each of those is an experiment, and each one is now public with weights attached.

The benchmark table is built to make one argument: a 6 billion active model can beat much larger ones on agentic work. All of the following numbers are vendor-reported. On SWE-bench Pro, Flash-Next scored 62.5 against 53.4 for Claude Opus 4.6 Max. On SWE-bench Multilingual it scored 81.0 against 77.5. On DeepSWE 1.1 it scored 58.7, with DeepSeek-V4-Flash-0731 the closest at 54.4. On CoWorkBench it scored 73.9 and on JobBench 55.7, the latter 19 points above the 36.6 Alibaba reported for Opus 4.6 Max. On GPQA Diamond it scored 91.7 and on LiveCodeBench v6 91.9. On Humanity's Last Exam without tools it scored 35.9, and that is the one language row where Alibaba shows Opus 4.6 Max ahead at 40.0.

Two caveats matter. First, the comparison target is Opus 4.6, not Opus 5, which shipped July 24 and now anchors Anthropic's lineup. Alibaba chose a model from two generations back for its headline comparison, and readers should weigh the table accordingly. Second, computer use is the visible gap. On OSWorld 2.0 the model scored 19.4 percent, level with the much smaller Qwen3.8-27B and far behind the 70.6 that Claude Opus 5 posts on the same test. Flash-Next is a coding and office-task model, not a desktop agent.

Alongside the weights, Alibaba announced a hosted production model called Qwen3.8-Flash on its QwenCloud API at $0.16 per million input tokens and $0.47 per million output tokens. The open weights ship under a custom qwen-community license, so check the terms before building a commercial product on them. The weights are on Hugging Face under the model ID Qwen/Qwen3.8-Flash-Next and on ModelScope.

Why this matters for practitioners: the 6B-active design means the model runs at roughly the inference cost of a 6 billion parameter dense model while carrying 125 billion parameters of knowledge plus a 51 billion parameter phrase table that lives in cheap memory. If the benchmark claims hold up under independent testing, that is a different cost curve for self-hosted agentic coding than anything available at the start of the summer. The Qwen4 family, when it arrives, will be built on the same ideas at larger scale.

DeepSeek added a vision model at Flash prices. On August 21 the company [listed deepseek-v4-flash-vision-exp](https://llm-stats.com/blog/research/deepseek-v4-flash-vision-exp-launch) on its pricing page and changelog. It accepts text plus images and returns text, with the same one million token context and 384K max output as V4-Flash and V4-Pro. Pricing matches V4-Flash exactly: $0.22 per million input tokens on cache miss, $0.007 on cache hit, and $0.66 output during off-peak hours, doubling to $0.44, $0.014, and $1.32 during the two peak windows at 01:00 to 04:00 and 06:00 to 10:00 UTC. Images are billed as input tokens at roughly 384 tokens per image after resize. Thinking is on by default with low, high, and max effort levels.

The self-reported text-agent numbers sit close to Flash-0731: Terminal Bench 2.1 at 83.9 against 82.7, DeepSWE at 59.3 against 54.4, NL2Repo at 57.7 against 54.2. The new vision-adjacent numbers are Chartography at 64.3 and ZeroBench Pass@5 at 35.0. DeepSeek said multimodal agents come close to Claude Opus 4.8, and Bloomberg reported the claim, but the changelog does not include an Opus column. Treat it as a sentence rather than a table. There are no confirmed open weights for this SKU yet. If you already run V4-Flash and have screenshot, chart, or UI loops that need image input, this is the same bill with vision added. Do not send images to the text V4-Flash or V4-Pro IDs, since they return a 400.

Two other model stories deserve a mention. A stealth model called Ox Alpha appeared on OpenRouter on August 20 under an anonymous provider, free to use, with a 1,048,576 token context and text, image, and video input. Bloomberg [reported](https://www.bloomberg.com/news/articles/2026-08-23/mystery-ai-model-ox-alpha-draws-developers-with-free-access) that developers rushed to it and that Stripe CEO Patrick Collison called it very impressive. Nobody has claimed it. Theories point at Zhipu, which has tested anonymously before, and at Microsoft's MAI family based on tokenizer analysis. The practical caution is simple: free inference from an unnamed party is a data policy you cannot read.

Zhipu's GLM-5.3, released August 14 through its coding subscription, has open weights due around August 28. The model shares its 744 billion parameter base with GLM-5.2 and gets its gains from post-training alone. Zhipu reports 84.5 percent on CyberGym, a cybersecurity capability benchmark, and cited that number as the reason for a longer safety review before publishing weights. Alibaba also finished open-sourcing Qwen3.8-Max, the 2.4 trillion parameter model with roughly 95 billion active, though the open checkpoint is text-only while the hosted version supports vision and a one million token window at $2 per million input and $6 per million output. And Alibaba's WAN 3.0 video model [launched August 24](https://llm-stats.com/blog/research/wan-3.0-launch), generating up to 30 seconds of 1080p video with audio in one pass, priced on fal at $0.05, $0.10, and $0.20 per second across tiers.

No new frontier flagship shipped this week from OpenAI, Anthropic, or Google. Gemini 3.7 Flash from August 13 is still the newest Google model, at $0.75 per million input and $3.75 per million output through the end of 2026, with the list price set to double on January 1, 2027. The Anthropic newsroom's most recent posts are from early August. OpenAI's news this week was about speed and price rather than a new model, and that belongs in the next section.

Two adoption notes round out the model picture. Moonshot's Kimi K3, the 2.8 trillion parameter mixture-of-experts model with 896 experts and 16 active per token, kept gaining commercial ground through August after its July 27 open-weight release. Legal technology company Harvey confirmed it built a new product on Kimi, which is one of the clearest signs yet of a Western enterprise shipping on a Chinese open model rather than only benchmarking one. Hosted Kimi K3 runs about $3 per million input tokens and $15 per million output, well above DeepSeek or Qwen pricing, because a model that size costs real money to serve even when the weights are free. Meta's Muse Code beta and Muse Spark 1.2 update from earlier in the month are still waiting on their promised open weights under a modified Llama Community License, while the 30 billion parameter Muse Glimmer is already ungated on Hugging Face under Apache 2.0. Llama 4 Behemoth remains unreleased more than a year after it was announced.

The week's biggest tooling story was a set of things becoming boring in the best way. On August 20, Anthropic announced that [computer use, the Skills API, and the Files API are generally available](https://claude.com/blog/computer-use-skills-api-files-api) on the Claude Platform, and shipped a new browser use tool inside computer use. The four pieces are designed to compose: an agent reads an intake document from the Files API, follows a skill that encodes a team's procedure, completes a form in a web portal with the browser tool, and saves the confirmation back as a file.

The computer use change that matters most is multi-action turns. The tool version computer_toolset_20260801 lets Claude take several actions per turn, click, type, key, screenshot, instead of one per round trip. Anthropic says early-access customers saw 20 to 40 percent fewer round trips per task, which shows up directly as lower latency and lower cost. Computer use is also now eligible for HIPAA-regulated workloads under Anthropic's business associate agreement, which opens it to healthcare automation that was previously off limits.

The browser use tool, browser_toolset_20260801, addresses the oldest problem in screen automation. Pixel-coordinate clicks break when a layout shifts. The new tool gives Claude the page structure alongside the screenshot, so it can act on element references rather than positions. NxCode's analysis makes an important operational point: both computer use and browser use are client toolsets. Claude proposes actions, and your application runs every click, keystroke, and navigation in an environment you control. That is the right security model, but it means you still own the executor, the credential isolation, the browser state, and the logic that halts a batch of actions when one fails.

The Skills API got simpler. You upload a folder of instructions, scripts, and templates once, version it, and pin requests to a specific version_id or to latest. The Files API now has five times higher rate limits and one terabyte of storage per organization, with automatic file expiration. The Skills API and Files API are available through Microsoft Foundry today, and Anthropic says the updated computer use and browser tools are coming soon to Google Cloud's Vertex AI. Existing beta integrations keep working during migration.

One data boundary detail is worth flagging for anyone in a regulated environment. Computer use and browser use can be zero-data-retention eligible on eligible models. The Files API and Agent Skills are not. If you are designing a workflow where some data cannot be retained, that asymmetry decides which pieces of the stack can touch it.

Anthropic also shipped Claude Academy, a free learning hub with courses and badges, and updated Claude Managed Agents so self-hosted sandbox sessions can attach memory stores, restrict web_search and web_fetch with allowed and blocked domain lists, and inspect multi-agent sessions in a redesigned console viewer. Add it all up and the message is that the agent building blocks are stable enough to build products on, and the remaining work is yours.

OpenAI's tooling news was about making its middle tier faster and cheaper. On August 18 the company [previewed an Ultrafast mode for GPT-5.6 Sol](https://openai.com/news/product-releases/) that it says runs up to 14 times faster than the model's standard speed, and cut GPT-5.6 Sol's API and credit pricing by more than 20 percent for three months. GPT-5.6 Sol is the model that Codex recommends by default, and it sits within half a point of Claude Opus 5 on Terminal-Bench 2.1 at 89.5 versus 89.1. The speed mode targets the places where only cheap fast models used to be viable: live voice, high-volume support, and coding assistants where a multi-second pause feels broken. The timing lines up with a summer of price pressure from DeepSeek, Qwen, and Zhipu. OpenAI is defending the tier developers reach for most often rather than only the top.

OpenAI also signaled a broader shift. Reporting on August 25 described the company scaling its agent strategy from specialized coding tools toward general-purpose consumer applications. The Codex team has spent a year building a harness for repository-level autonomy. The next step is pointing that harness at everything else.

The harness question got a fresh answer from research. The Laude Institute open-sourced Headlong, an agent harness under 10,000 lines of Bash that keeps a model in a continuous self-guided inner-monologue loop rather than the request-response pattern most frameworks use. In demos an agent named Audel debugged its own code and started projects with no human prompt. The loop runs at roughly $1 to $2 per hour with exponential backoff when idle. It is a research artifact, not a product, but it is a clean example of the persistent-agent pattern that Claude Code, Codex, and Cursor are all edging toward with background agents and scheduled tasks.

In the enterprise tooling lane, Glean unveiled Glean Tau on August 26, a desktop workspace that connects its enterprise search and agents to a user's local files, applications, and code. Glean claimed a token-cost edge over Claude, which is a claim to verify rather than repeat. And CellCog's August rankings of agent harnesses put Claude Code first for depth of hooks, subagents, and workflow control, with Codex CLI highlighted for cloud-based pull-request-shaped autonomy, Cursor leading in-editor agent workflows, and Gemini CLI and GitHub Copilot rounding out the top five.

Two surveys give the human side of the picture. A Coddy developer survey covered by ZDNet found that 80 percent of developers describe their AI coding tool usage as feeling more like dependence than advantage, citing the loss of natural stopping points like waiting on a review. LeadDev's 2026 leadership survey found 45 percent of engineers work more hours per week than the year before. And Reuters reported that Meta wanted to replace far more of its workforce with AI agents than previously known, and that the plan collapsed under employee pushback and agents that failed to deliver. The tools are getting better every month. The organizational questions are not getting easier.

The standards story of the week is a change of address. Axios [reported on August 17](https://www.axios.com/2026/08/17/a2a-agentic-ai-foundation-open-ai-standards) that the Agent2Agent Protocol, the Google-created standard for agents to talk to one another, is moving from the Linux Foundation's broader portfolio into the Agentic AI Foundation as a hosted project. That puts A2A in the same home as the Model Context Protocol, which handles connections between an agent and its tools and data. AAIF launched in December 2025 with fewer than 40 members and now counts more than 250, including Google, Microsoft, Amazon, Anthropic, OpenAI, Bloomberg, Shopify, and Block.

For anyone who has not followed the protocol stack, the split is simple. MCP is vertical: it connects one agent to a database, a file system, an API, or a catalog. A2A is horizontal: it lets one agent hand a task to another agent without either exposing internal state. An A2A agent publishes an Agent Card at a well-known URL describing what it does and how to authenticate, reusing OpenAPI security schemes for API keys, OAuth 2, OpenID Connect, and mutual TLS. Signed agent cards let a caller verify the card has not been tampered with, which matters because a poisoned card can redirect everything that trusts it. The third contender, IBM's Agent Communication Protocol, folded into A2A in 2025, so there is one agent-to-agent standard worth building against.

AAIF executive director Mazin Gilbert framed the move in terms of the whole stack. Companies do not want just one open protocol, he told Axios. They want the entire stack to be open and interoperable. Google Cloud VP Rao Surapaneni said the original A2A hypothesis was that customers deploy agents from multiple providers and all of those agents need to work together. A2A already ships natively in Azure AI Foundry, Amazon Bedrock AgentCore, and Google Cloud, with more than 150 organizations supporting it as of April. The governance change does not alter the protocol, but it does put MCP and A2A under one roof at the moment both are stabilizing, and it makes the reported joint MCP and A2A specification effort easier to run.

The MCP side of that roof is settling into its new shape. The 2026-07-28 specification, released a month ago, replaced the session-based protocol with a stateless core. The initialize handshake and Mcp-Session-Id are gone, every request is self-describing through _meta and HTTP headers, and servers can deploy on serverless and edge infrastructure behind a plain round-robin load balancer. Three official extensions ship under a versioned framework: MCP Apps for server-rendered UI, Tasks for long-running operations, and Enterprise Managed Auth for IdP-based provisioning. Roots, Sampling, Logging, the HTTP plus SSE transport, and Dynamic Client Registration are deprecated with a 12-month removal window. This week's news is adoption rather than change. Anthropic's connector directory has passed 950 servers. Cloudflare's Agents SDK supported the spec from day zero. AWS shipped the stateless core in Bedrock AgentCore. Supabase said the new multi-round-trip request mechanism finally lets its stateless server ask a user to confirm before deleting data.

Simon Willison's take a few weeks ago captured the mood: MCP had been eclipsed by Skills once it became clear that an agent with a terminal and curl can do most of what MCP did, and the stateless redesign gave the protocol a clearer job. Skills teach an agent how to use existing software. MCP gives many clients a shared contract for discovering and calling a remote tool with authorization and discovery built in. The two compose. A skill can describe when to use an MCP server and what its domain concepts mean, and MCP can handle the remote execution boundary.

That framing turns Anthropic's Skills API general availability into a standards story too. A skill is a folder with a SKILL.md file, scripts, and templates. The format is open and file-based, which is why third-party tools already expose skill folders through MCP servers so any MCP client can load them. With the Skills API now in production, versioned, and available through Microsoft Foundry, the SKILL.md format is becoming the de facto standard for packaging agent procedures the way MCP became the standard for packaging tools. Nobody has written a formal spec for it. That is usually the step right before someone does.

One more standards-adjacent item. Google Cloud published guidance on August 24, tied to its State of AI Infrastructure report, that frames agent security as the top gating issue for scaling autonomous workflows. The recommendations are platform-level governance, task-level provenance, per-task permissions, and human-in-the-loop checkpoints with end-to-end audit trails. None of that is a standard yet. All of it is what MCP's Enterprise Managed Auth extension and A2A's signed agent cards are reaching toward, and the fact that a hyperscaler is writing the playbook is a sign the protocols will be asked to carry it.

Hot Chips 2026 ran August 23 to 25 at Stanford, and the theme across Nvidia, Google, OpenAI, Microsoft, Meta, AMD, and SambaNova was the same: agents generate enormous token volumes across hundreds of inference steps, and the generation phase, where a model emits one token at a time, is where responsiveness and cost are decided. Three presentations stood out.

Nvidia [announced on August 24](https://nvidianews.nvidia.com/news/nvidia-groq-3-lpx-now-in-full-production-with-world-class-speed-for-agentic-ai) that Groq 3 LPX, the inference accelerator built from technology it acquired in its $20 billion Groq deal last December, is in full production. LPX is a rack-scale system designed to sit alongside Vera Rubin NVL72. Rubin GPUs handle prefill and heavy context processing, and the language processing units handle the latency-sensitive decode phase. In Artificial Analysis benchmarking on Gemma 4 31B with a 100,000 token context, the system delivered 3,400 output tokens per second, which Nvidia says is four times faster than the nearest alternative platform. StorageReview reported the rack holds 256 LP30 chips. Nebius is the first AI cloud to adopt LPX, planning to serve it through Nebius Token Factory behind the same API developers already use. CoreWeave and SpaceXAI were named as early Vera Rubin platform adopters, and Groq itself plans to be among the earliest LPX customers.

Nvidia senior director Dion Harris was careful to say the chip is not meant to replace the GPUs that train and run most workloads, only to handle the low-latency slice of inference. That is a notable admission. Nvidia is now shipping a non-GPU architecture for a phase of inference where its GPUs were not winning on latency, and it bought the company that was. The company also says agentic workloads consume roughly 15 times more tokens than a simple chat request, and it projects a combined $1 trillion in sales from Blackwell and Vera Rubin through 2027.

Google [presented its eighth-generation TPU family](https://www.servethehome.com/googles-tpuv8s-for-training-and-inference-at-hot-chips-2026/), and the headline is that it built two chips in one year instead of alternating. TPU 8t is for training and TPU 8i is for inference. Google made the case with a pop quiz: shown two chips, one smaller with six HBM stacks and one larger with eight, most of the audience guessed the smaller one was for inference. It was the training chip. Inference needs more HBM per unit of compute and a higher share of SRAM, because the cost of low-latency serving rises steeply as per-user token rates climb. The 8i pairs with Google's own Axion Arm CPUs at a 2-to-1 ratio, replacing x86, uses a new BoardFly network topology with a maximum of 7 hops instead of 16 for the old 3D torus, and moves collective operations into the I/O die so they never touch the compute die or HBM.

The 8t numbers are large. A superpod holds 9,600 chips across 300 racks with access to 2 petabytes of shared HBM and 121 exaflops of FP4 compute, at roughly twice the performance per watt of TPUv7 Ironwood. A new dedicated Virgo network supports 134,000 TPUs in a single domain at 47 petabits per second. Google is liquid cooling its optics for the first time, runs in-field unit tests during idle cycles to catch failing chips, and used AI models to trim TPU 8t power and area during design. Existing code runs unchanged, and Google's Pallas kernel language gives hardware-aware Python for both chips.

OpenAI closed the conference with [a deep dive on Jalapeño](https://www.servethehome.com/openai-jalapeno-asic-at-hot-chips-2026/), the in-house inference ASIC it built with Broadcom in roughly nine months from initial RTL to tapeout. The spec sheet reads 13.4 petaflops of MXFP4 compute, 15.4 terabytes per second of HBM4 bandwidth across 216 GiB, and a 700 watt package, scaling to 27 exaflops and 432 TiB across a 2,048-chip system. A local domain of 128 chips gets 600 gigabytes per second of interconnect, and a global domain of 2,048 chips built on Broadcom Tomahawk6 switches gets 200. The design timeline shows an architecture concept in late 2024, RTL freeze in 2025, a late 2025 tapeout, Codex running on the chip in early 2026, and ChatGPT following soon after.

OpenAI's framing is the interesting part. It measures two things: time to last token for user experience and tokens per joule for cost. It benchmarks on InferenceX, a public power-normalized suite, with Jalapeño at 700 watts against GB200 at 1.2 kilowatts and GB300 and MI355X at 1.4 kilowatts. On GPT-OSS 120B it claims about 1.9 times higher peak mixed tokens per second per kilowatt and 1.7 times lower end-to-end latency at matched operating points. On DeepSeek R1 at 670 billion parameters it claims 1.7 times and 3.6 times. On the one trillion parameter Kimi K2.5 it claims 1.5 times and 3.4 times. The comparisons use single-token prediction on Jalapeño against multi-token prediction on the Nvidia baselines, and OpenAI says adding MTP on its own chip adds another 3 to 5 times latency improvement. All of these are OpenAI's numbers on OpenAI's chosen benchmark. Independent validation does not exist yet.

The architectural argument deserves attention regardless. OpenAI says a single request spans three regimes: compute-bound prefill, a tiny draft model at ultra-low batch, and memory-bound speculative-verify decode with bursty MoE communication. Rather than a heterogeneous fleet where each phase runs on specialized silicon and the KV cache travels between them, Jalapeño keeps KV local and varies which units are active per phase, gating the idle blocks. Each core slice is paired with its own HBM slice for a fast local view. The programming model, called Gluon, treats each physical core as a thread block with explicit tensor placement, built so an AI search can handle the mapping and scheduling. OpenAI said its internal model drove attention and MoE kernels to 1.5 to 1.8 times the speed of expert-written implementations, and that AI-assisted design found a 56 percent improvement on a BF16 multiply with a 10 percent smaller matrix unit. This is Gen 1 of a multi-generation roadmap.

Taken together, the three presentations describe one design pressure. Groq 3 LPX is a decode specialist bolted onto a GPU rack. TPU 8i is an inference chip with more memory and SRAM than its training sibling. Jalapeño is a balanced chip that gates itself between phases rather than moving data between chips. Three different answers, one question: how do you serve a trillion parameter MoE model to one user at low latency without wasting the rest of the rack?

The answer is going to cost more than expected. Bloomberg reported on August 24 that Nvidia has told its largest customers that servers containing its chips will rise more than 15 percent in price in many cases for systems shipping early next year, including Vera Rubin and Grace Blackwell configurations. The driver is not the GPUs. It is memory. Samsung, SK Hynix, and Micron produce most of the world's high-bandwidth memory, and their output has not kept pace with demand even after ramping through the year. HBM4 is expected to pass 50 percent of HBM sales in the second half of 2026 at prices 60 to 70 percent above HBM3E. Every chip at Hot Chips was designed around more HBM per unit of compute, and every one of them competes for the same three suppliers' output. The Hot Chips memory tutorial day featured Micron, Samsung, SK Hynix, d-Matrix on 3D DRAM, and Oxmiq Labs on high-bandwidth flash, which tells you where the industry thinks the bottleneck is.

Nvidia reports quarterly earnings on August 26, and the price increases land the same week. Analysts expect another large quarter with growth rates that decelerate against a much bigger base. For anyone budgeting AI infrastructure into 2027, the practical message is that the cost of building capacity is rising even as the price of using hosted models keeps falling.

Every one of this week's stories has a data implication, and it is worth stating plainly.

The Qwen3.8-Flash-Next N-gram embedding layer is a 51 billion parameter lookup table that sits in system RAM. That is a data structure, not a neural network, and it is the first time a major open-weight model has shipped a large chunk of its capacity in a form that looks more like a key-value store than a tensor. Expect inference engines to treat it like one, with the same caching and tiering tricks used for KV caches and embedding tables.

DeepSeek's vision SKU at Flash prices makes chart and screenshot reading cheap enough to run on every dashboard, every report, and every UI regression test. The bottleneck for that workload moves from model cost to getting the images and the context to the model, which is a data pipeline problem.

Anthropic's Files API at one terabyte per organization and the Skills API in production mean agents now have durable storage and durable procedures. The next question is which of those files and skills should live in a governed catalog with lineage, versioning, and access control, and the answer is most of them.

A2A's signed agent cards and MCP's Enterprise Managed Auth are both answers to the question of which agent is allowed to touch which data. The catalog layer is where those permissions will end up being enforced, because it is the only place that already knows what the data is.

And Hot Chips made the case that the token-generation phase is where inference economics are decided. The agent loop is inspect, plan, act, verify, repeat. Every step of that loop that touches a table or a document is a query. Fast decode makes the model's part of the loop faster. It does nothing for the query, which means the data layer's latency is about to be the visible part of the agent's latency.

Here is what to do with all of this if you build or run AI systems for a living.

Test Qwen3.8-Flash-Next on your own coding tasks before believing the table. The vendor numbers are strong, the comparison target is two generations old, and the model is weak on desktop control. The right evaluation is your repository, your issue tracker, and your CI. If it holds up, the 6B-active cost profile changes the math on self-hosted agents. If it does not, you have lost an afternoon. Either way, read the architecture notes, because the N-gram embedding layer and Qwen Sparse Attention are the shape of Qwen4.

Route image-heavy loops to the cheapest capable vision model, and for many teams that is now DeepSeek's Flash vision SKU at $0.22 per million input tokens off-peak. Budget the peak windows. Set detail to low when you do not need fine pixels, since images cap at roughly 384 tokens each after resize. Keep the experimental label in mind and do not build a hard dependency until a GA identifier exists.

If you are on the Claude Platform, migrate to the 20260801 computer use and browser toolsets now rather than waiting for the beta identifiers to be removed. The multi-action turns cut round trips by 20 to 40 percent in early access, which shows up in your bill. Build the executor with ordered actions and a halt-on-failure path, and put approval checks in front of any consequential step. Put your team's procedures into versioned skills and pin production requests to a version_id rather than latest, so a skill edit cannot change production behavior without a deploy.

Check the data retention boundary before you route regulated data. Computer use and browser use can be zero-data-retention eligible. The Files API and Agent Skills are not. Design the workflow so regulated content flows through the tools that support your policy and never lands in a file or skill that does not.

If you serve GPT-5.6 Sol through Codex or the API, try Ultrafast mode on your latency-sensitive paths during the three-month price cut. Voice, support, and inline coding assistance are the obvious candidates. Measure quality alongside speed, because a mode that runs 14 times faster is worth exactly nothing if it produces worse patches.

Do not send production prompts to Ox Alpha. Free, anonymous, one million tokens of context, and a promise not to train on your data is a promise from nobody. Use it for curiosity and public benchmarks. Wait for a name before you wire it into anything.

If you are building multi-agent systems, standardize on A2A for agent-to-agent handoffs and MCP for tool access, and stop evaluating alternatives. ACP is gone. ANP is not production-ready. Both surviving protocols now live under the same foundation, the joint specification work is coming, and every major cloud has native support. Sign your agent cards. Use the OpenAPI security schemes you already understand. Treat the Enterprise Managed Auth extension in MCP as the direction of travel for how agents get org-wide identity.

Migrate MCP servers to the 2026-07-28 spec on your own schedule but inside the 12-month deprecation window. Servers that relied on sessions, handshake-time configuration, or connection-local state need real review. Simple tool servers mostly need an SDK bump. Stateless servers can move to serverless and edge hosting behind an ordinary load balancer, which for many teams is the first time an MCP deployment has been cheap to run.

If you are packaging agent procedures, write them as SKILL.md folders even if you do not use Claude. The format is plain files, it loads through MCP servers into any client, and it is on its way to being the common currency for agent know-how the way MCP became the common currency for tools.

Budget more for hardware next year, not less. The memory shortage is real, Nvidia is passing it through, and HBM4 pricing is heading up rather than down. If you were planning capacity for 2027 on the assumption that hardware costs keep falling with model prices, revisit the plan. Hosted inference keeps getting cheaper because the labs are absorbing the hardware cost to compete. That is a subsidy, and subsidies end.

Watch the decode phase in your own latency budgets. The chips at Hot Chips all attack token generation because that is where user-visible latency lives. Your agent loop's latency is model time plus tool time plus data time. As the model time shrinks, the tool and data time become the part users notice. If your agent queries a table on every step, the query planner is now on the critical path.

And read the developer surveys with your own team in mind. Eighty percent of developers describing their tool use as dependence, and nearly half working more hours, is not a tooling problem that a better model fixes. It is a process problem. Agents remove the natural pauses in a workday. Put some back on purpose.

GLM-5.3 weights are due on Hugging Face around August 28, which will give the community an independent look at Zhipu's cybersecurity numbers. Nvidia's earnings on August 26 will put financial numbers behind the Hot Chips claims. Qwen3.8-Flash-Next will get its first independent benchmark runs, and the OSWorld gap is the thing to watch. Anthropic's IPO reporting is intensifying, with the Financial Times describing investor targets of a $2 trillion valuation as early as October and a revenue run rate of $65 billion by the end of July. Ox Alpha's owner will be unmasked eventually, and the tokenizer analysts think they already know. And the first A2A project meetings under the Agentic AI Foundation should set the agenda for the joint MCP and A2A specification work.

If you want to go deeper on how agents, models, and the data layer fit together, I write books on agentic AI, AI-assisted development, Apache Iceberg, and the lakehouse. You can find all of them at [books.alexmerced.com](https://books.alexmerced.com).