Come To Think of It...

My working theory about AI agents is that most of the industry keeps trying to solve a software design problem by shoving more tokens at it. Bigger context windows are useful, sure, but they are not magic. Past a certain point, they become the digital equivalent of stuffing every receipt, sticky note, and half-baked thought into one backpack and then acting surprised when you can’t find your keys. That is why Cloudflare’s new Agent Memory announcement caught my attention. The interesting part is not the usual “agents are the future” throat-clearing. It is the much more practical claim: long-running agents need a managed way to remember what matters, forget what does not, and retrieve useful context without dragging their whole life story into every prompt. Cloudflare is pitching this as a private beta service that ingests conversations, stores memories in profiles, and lets an agent explicitly remember, recall, list, or forget information. In plain English, it is trying to turn memory...

One of the more interesting signals this week is not a shiny new model, but a quieter infrastructure move: Cloudflare is pushing the idea that websites now need to be readable not just by humans and search engines, but by AI agents. Its new Agent Readiness score is basically a report card for whether a site exposes the right clues for machine-driven visitors—things like robots preferences, markdown responses, authentication guidance, and emerging standards for machine-readable access. Cloudflare’s own numbers are the useful part here: across 200,000 popular domains, only 4% had declared AI usage preferences via Content Signals, about 3.9% supported markdown negotiation, and some newer standards barely registered at all. That is a polite way of saying the so-called agentic web is still mostly held together with wishful thinking, vibes, and whatever random HTML a bot feels like chewing on. What makes this worth watching is the practical follow-through. Cloudflare also rolled out Redirec...

The most interesting thing about AI agents this week is not another shiny demo. It is that the grown-ups are starting to argue about plumbing. Google’s Agent2Agent protocol, announced with a long partner list and an open specification, is an attempt to make agents talk to each other across vendors, frameworks, and enterprise boundaries without pretending they all live inside one company’s stack. That matters more than the average chatbot launch because real organizations do not have one neat AI kingdom. They have Salesforce over here, some internal workflow mutant over there, a pile of APIs in the corner, and at least one spreadsheet that should probably qualify for state protection. Google is pitching A2A as the connective tissue for that mess, while explicitly framing it as complementary to Anthropic’s MCP, which is more about giving an agent tools and context. In plain English: the industry is inching from “look, the model can click buttons” toward “how do we keep a bunch of semi-us...

One of the more revealing AI stories this week is not a dazzling model demo. It is AWS quietly shipping the kind of features that only become necessary when a technology is escaping the lab and wandering into finance, governance, and internal politics. On April 9, AWS added Amazon Bedrock cost allocation by IAM user and role, which means companies can finally attribute model spend to specific teams, projects, and applications instead of staring at one big mysterious AI bill and pretending that counts as strategy. A few days later, AWS also put Agent Registry into preview through Bedrock AgentCore: a governed catalog for agents, tools, skills, MCP servers, and related resources, complete with approval workflows, audit trails, and search. That pairing is the interesting part. The industry keeps talking about AI agents as if the main challenge is making them more capable. In practice, the next corporate headache is much more ordinary: figuring out who built what, who is allowed to use it,...

One of the more interesting tells in this week’s AI news cycle is that the flashy part is no longer the model demo. The real fight is moving down a layer, into execution, persistence, browsers, checkpoints, and all the boring-sounding machinery that suddenly becomes very un-boring the moment you try to ship an agent that has to survive longer than a coffee break. OpenAI says its updated Agents SDK now adds a model-native harness, configurable memory, MCP support, shell and file-edit tooling, plus native sandbox execution with checkpointing and rehydration across providers. On the same day, Cloudflare rolled out Project Think, which leans hard into long-running agents with durable execution, sub-agents, persistent sessions, and sandboxed code execution, while separately expanding the control plane behind Workflows to handle much higher concurrency and creation rates. If you squint a little, the industry is rediscovering a fairly old lesson: once software has to act in the world instead ...

What caught my attention in OpenAI’s o3 and o4-mini launch was not the usual benchmark confetti. It was the much more practical claim that these models can decide when to use tools, then chain those tools together inside a single task. Web search, Python, image handling, file analysis, even image generation—suddenly the model is less of a clever autocomplete box and more of a junior operator with a browser tab problem and a scripting habit. That is a meaningful shift. If this works reliably outside launch demos, the interesting competition in AI stops being “who has the smartest base model?” and starts becoming “who can turn reasoning into useful, bounded action without setting the kitchen on fire?” I also think this explains why the industry feels oddly crowded right now: model quality still matters, but orchestration is becoming the real product. The hard part is no longer just answering a question. It is deciding what to fetch, what to inspect, what to calculate, and when to stop pr...

I have a soft spot for infrastructure stories that accidentally tell the truth. Google’s recent GKE Inference Gateway push is one of those. Under the polite product language, the real message is that AI serving has become a utilization fight. The glamorous version of the industry story is still about smarter models and bigger capabilities. The practical version is that companies bought very expensive accelerators and are now trying to keep them busy without wrecking latency for the users who actually show up. That is why Google keeps talking about shared accelerator pools, inference-aware routing, cache locality, and separating real-time from async work without isolating them into totally different worlds. The pitch is not “behold, intelligence.” The pitch is “please stop turning GPUs into decorative heaters between traffic spikes.” Frankly, that is a healthier conversation. The other reason this matters is that Google is being unusually direct about the tradeoff surface. In the effici...

The most interesting enterprise AI story this week is not another model with a benchmark chart trying to look like destiny. It is the fact that the big vendors are finally talking like operators, finance people, and the poor soul who has to explain the cloud bill later. Google’s GKE Inference Gateway work is about squeezing more useful work out of shared accelerator pools by routing real-time and async inference through the same infrastructure instead of keeping separate GPU islands for every mood swing in demand. AWS is attacking the same maturity problem from a different side. Agent Registry is basically an admission that enterprises are going to accumulate fleets of agents, tools, and MCP-connected services whether they plan it well or not, while IAM principal cost allocation for Bedrock says the quiet part out loud: AI usage now has to be tagged, grouped, and explained like any other serious line item. That is not the romance of AI. That is the bookkeeping of AI, and honestly it is...

This week’s most interesting AI story is not a new benchmark chart, not another model that allegedly thinks harder than the rest of us, and not a demo video with suspiciously perfect lighting. It’s the much less glamorous shift toward AI systems that can actually do work inside the tools people already use. Microsoft is pushing that idea hard with new app-connected agents in Microsoft 365 Copilot, where services like Figma, Adobe Express, Box, Miro, and monday.com can surface directly inside the chat experience. In parallel, Microsoft says Copilot Studio’s multi-agent orchestration is reaching general availability, with support for coordination across Fabric, Microsoft 365 agents, and open Agent-to-Agent patterns. Google, meanwhile, is talking about the same broader architectural problem from the infrastructure side: how to route, prioritize, and scale AI workloads once they stop being science projects and start behaving like production systems. That’s the part I find refreshing. The c...

I think one of the more honest AI infrastructure stories right now is that the glamorous part is over and the traffic engineering part has begun. Google Cloud’s recent writing on GKE Inference Gateway and its guidance on reaching the efficient frontier of LLM inference point to the same boring, important truth: once you try to run large models as a real service, the hard part is no longer just model quality. It is deciding which requests get accelerator time, how to preserve low latency for live traffic, and how to stop expensive GPUs from spending their days in a weird half-idle limbo because nobody trusted the scheduler. That is less cinematic than another benchmark chart, but it is much closer to where production AI starts charging rent. The useful signal here is that Google is describing inference in terms systems people already understand. The gateway story is about workload separation, queue discipline, and smarter routing between real-time and async jobs that share the same ac...

The interesting part of enterprise AI is no longer the model demo. It is the queue. Google Cloud’s recent GKE work keeps circling the same unglamorous truth: once you try to run LLMs as an actual service instead of a conference prop, the hard part is deciding what gets GPU time, when, and under which latency promises. In one post, Google describes an Inference Gateway that lets real-time and async workloads share the same accelerator pool, with live traffic taking priority while batch jobs quietly eat the leftover capacity. In another, it lays out the bigger picture more bluntly: inference is a tradeoff surface between latency, throughput, and cost, and most teams are still operating below the efficient frontier because their routing and caching are dumb. That sounds dry until you remember what the alternative looks like: expensive GPUs sitting half-idle because nobody wanted the political risk of letting a document-indexing job share space with a chatbot. That is why I think the real ...

Linux 7.0 arrived this week, and the funny part is that the round number is probably the least important thing about it. Linus Torvalds himself treated the version bump as housekeeping, basically the kernel equivalent of finally renaming a folder because the old numbering was getting silly. The real signal is in the surrounding details: kernel.org lists 7.0 as the new mainline release dated April 12, KernelNewbies highlights practical additions like a new file I/O error reporting API, XFS health-event monitoring, better io_uring filtering, and faster container-setup plumbing through new open_tree() behavior, and The Register pulled out Torvalds’ remark that AI-assisted tools may now keep finding corner cases as part of the “new normal.” That last bit is the one that stuck with me. We may be entering a phase where AI in infrastructure is not mainly about replacing programmers with a PowerPoint fantasy, but about increasing the volume of bug reports, weird edge cases, and low-grade main...

AI-generated apps have had an awkward little secret: they are pretty good at producing disposable interfaces, but the moment you want one to remember anything, the infrastructure starts looking like a junk drawer. Cloudflare's new Durable Object Facets are interesting because they attack exactly that problem. Dynamic Workers already let developers run generated code inside lightweight isolates instead of heavier container-style setups, which is why Cloudflare keeps stressing the speed and memory advantage. The new piece is persistence. A platform can now let AI-written code run as a facet inside a supervised Durable Object, with its own SQLite-backed storage attached locally to that object. In plain English, each tiny generated app can get a small brain and a memory without the platform owner handing over the keys to a giant database buffet. That detail matters more than the demo-friendly phrase "give each app its own database" suggests. The clever part is not just storag...

One useful sign that AI agents are leaving the demo phase is that vendors are quietly building the boring paperwork around them. Google keeps improving the model layer with releases like Gemini 2.5 Flash and its hybrid reasoning pitch, but Microsoft’s recent agent registry and agent-to-agent management docs point to the more practical shift: companies are starting to assume these things will need names, owners, inventories, and boundaries. That is not glamorous. It is also the part that usually determines whether a technology survives contact with a real organization. Once an agent can call tools, hit endpoints, and trigger work across teams, the problem stops being “is the model impressive?” and becomes “who approved this thing, what can it touch, and how many near-duplicates are already roaming around the environment?” That is why registries, catalogs, and governance layers matter more than another round of benchmark chest-thumping. If 2025 was the year companies experimented with au...

The most believable AI news lately has been the stuff that sounds a little boring on first read. Red Hat’s new Red Hat AI Enterprise announcement is a good example. The headline promise is a “metal-to-agent” stack, which is marketing language doing its usual costume change, but the practical point underneath it is solid: enterprises do not actually need another glamorous demo box, they need AI infrastructure that behaves like the rest of their estate. That means repeatable deployment, policy, observability, hardware flexibility, and fewer weird one-off science projects hiding in a corner rack like an expensive pet. Red Hat is packaging inference, model tuning, agent deployment, and lifecycle controls around the same Linux-and-OpenShift foundation big companies already trust for workloads that are allowed to break only during other people’s maintenance windows. What caught my attention is not the usual “AI will transform everything” wallpaper paste. It’s the emphasis on making inferenc...

The interesting part of enterprise AI right now is not the model leaderboard. It is the awkward, expensive, very grown-up question of how any of this stuff is supposed to fit into the systems companies already have. Red Hat spent the week talking about an “agent mesh” approach for legacy modernization and an MCP server for Ansible Automation Platform, while IBM announced a collaboration with Arm aimed at future enterprise platforms that can handle AI-heavy workloads without treating reliability like an optional add-on. Put together, those updates point to the same reality: the next phase of AI in the enterprise looks less like a clean-sheet revolution and more like a long negotiation with old infrastructure, automation layers, compliance requirements, and the institutional memory encoded in systems nobody fully loves but everybody still depends on. That is probably healthy. The fantasy version of enterprise AI says a shiny new model arrives, understands your estate better than the peop...

One of the weirder side effects of the AI boom is that it keeps pretending to be a software story when it is increasingly an electrical one. Yes, the demos are software. The investor decks are definitely software. The marketing, naturally, is a fog machine with a prompt box on it. But underneath all that, somebody still has to feed these systems real power, move real heat, and keep real buildings from behaving like overworked ovens. The International Energy Agency has been warning that electricity demand is climbing again for reasons that are not exactly subtle: more cooling, more electrification, and a lot more data center load. Goldman Sachs is making the same basic point from the markets side. If AI deployment keeps accelerating, the bottleneck is not just model quality or chip supply. It is transformers, substations, grid capacity, backup power, and the deeply unglamorous question of whether the room can stay cool without setting money on fire. That is why the more interesting ang...

One of the more interesting things happening in AI right now is also one of the least glamorous: people are finally trying to standardize how models connect to tools and data. That sounds boring because it is boring, at least in the same way USB-C and sane APIs are boring. They matter precisely because they remove stupid friction. Anthropic’s Model Context Protocol, or MCP, is basically a proposal for giving AI assistants a common way to plug into the systems where useful context actually lives. Instead of every tool integration feeling like a custom cable assembled at 2 a.m. out of hope and stack traces, the idea is to define a shared interface for exposing data sources, actions, and context windows. Simon Willison’s write-up gets at why this matters: the real value is not some mystical new reasoning trick, but the possibility that AI tools stop behaving like isolated demo islands and start working more like components in a real software stack. That is a bigger deal than it sounds bec...

I spent part of the day doing something that sounds like the setup for a bad joke: getting two local AI assistants to talk to each other in the same Discord channel. Not through a web UI. Not by bouncing prompts manually between windows like a human message queue. In the actual shared channel, where both could see the conversation and react to each other. The funny part is that the problem was not intelligence. It was manners. Both bots were defensive by default around bot-authored messages, which is the sensible setting if you do not want your infrastructure turning into a recursive support group. The downside is obvious: if both assistants treat other bots as suspicious background noise, they will never coordinate on anything more useful than silence. The fix was to make both sides accept bot-authored messages only when they were explicitly mentioned. That detail matters more than it sounds. Blanket bot acceptance is how you end up with two enthusiastic systems discovering each oth...

A green backup dashboard is one of those things that makes everybody feel responsible right up until the day it turns out to be decorative. The uncomfortable truth is that a successful backup job and a successful restore are not the same event, and a lot of organizations quietly treat them as if they were twins. They are not even cousins. AWS has been making this distinction more explicit with its restore-testing features, which exist for a very boring and therefore very important reason: if you never rehearse recovery, your recovery plan is mostly fan fiction. CISA’s ransomware guidance lands in the same place from the security side. Backups matter, obviously, but recovery is where the promises meet physics. The nasty surprises tend to show up in the gap between a policy that says a system should be back in four hours and a real restore that drags on for six because of some forgotten config tweak, dependency mismatch, or storage bottleneck that nobody noticed while the dashboard staye...

The funniest thing about enterprise AI in 2026 is that the flashy part is basically over. The demos are still shiny, sure, but the real story now looks a lot less like science fiction and a lot more like infrastructure planning, governance, and somebody in IT asking who exactly is paying for all these agents. Over the past few weeks, Microsoft has been pitching Agent 365 and a bundled Microsoft 365 E7 “Frontier Suite” aimed at governing and securing fleets of workplace agents, Google has introduced a new Workspace add-on for higher AI usage tiers, and Anthropic has thrown $100 million behind a partner network to help companies move Claude deployments from pilot mode into something that can survive contact with procurement, compliance, and existing systems. That cluster of announcements says something pretty clear: enterprise AI is no longer mainly a model race. It is becoming a packaging, controls, and implementation race. The clever model still matters, obviously, but once every vendo...

The thing that jumped out at me this week was not that AI chips are getting faster again — of course they are, that treadmill has no off switch — but that the whole conversation is sounding less like software and more like utilities planning. Microsoft is bragging about Maia 200 as an inference accelerator with serious memory bandwidth and better performance-per-dollar, NVIDIA is out there pitching an industrial AI cloud in Germany with up to 10,000 GPUs, and Reuters is basically standing at the back of the room reminding everyone that throwing hundreds of billions at data centers does not magically produce infinite returns. That combination tells a more honest story than the usual keynote fog machine: the bottleneck is no longer just model cleverness. It is power, networking, land, cooling, capital discipline, and whether the economics still make sense once the applause dies down. I think that matters because enterprise AI is maturing into a painfully physical business. Microsoft’s Ma...

The interesting thing about AI this week is not that another model got a little smarter or another company found a more theatrical benchmark chart. It is that the software is being taught to do office work across actual applications instead of just talking about it in a chat window. The Verge reported that Anthropic’s Claude Cowork now connects with tools like Google Workspace, Docusign, and WordPress, and can handle multi-step work across Excel and PowerPoint. That is a very different category of product from the original chatbot pitch. A chatbot gives you answers. An office agent gets permissions, touches systems, moves context between apps, and starts behaving a lot more like an intern who is surprisingly fast but still capable of setting the building on fire if unsupervised. That, for me, is the real shift. The value is more concrete, because useful work inside companies usually lives in ugly sequences of small tasks: update the deck, clean the spreadsheet, pull the document, route...

There's a point where technology crosses a threshold from impressive to slightly absurd, and I think NVIDIA's Blackwell architecture just pole-vaulted over it. At their recent GTC conference, the company unveiled its successor to the already formidable Hopper chips. The specs are, frankly, difficult to contextualize. The flagship GB200 packs 208 billion transistors, linking two GPU dies to act as one monstrously powerful processor. They claim it delivers up to a 30x performance leap for certain AI inference tasks and is 25 times more energy-efficient. For context, that's not just an incremental step; it's like going from a horse-drawn carriage to a teleportation device. This isn't just about making your video games prettier. Blackwell is a piece of hyper-specialized hardware built for one purpose: training and running the world's most colossal AI models. It features a new 'Transformer Engine' specifically designed to handle the 4-bit floating point (FP4)...