Delete your CLAUDE.md. Not edit it, not trim it, delete the whole file, along with your skills and your hooks, and see what the model does without them. That's Boris Cherny's actual advice from a recent YC appearance, and the Claude Code team has apparently followed it internally enough times that roughly 80% of their own system prompts are gone. His reasoning: most of those rules patched a weakness in whatever model you ran six or twelve months ago. Newer models often don't have that weakness anymore, so the instruction that used to help is now just constraining a more capable model for no reason. His method is almost embarrassingly plain. Delete everything, work normally for a while, watch where the model actually goes off track, add back exactly one instruction when a mistake repeats, and run the whole cycle again at the next major model release. Keep security and compliance rules. Keep the project quirks a model genuinely can't infer from reading the code. Cut folder structures, dependency lists, and architecture summaries, since a capable model can read those off the repo itself faster than it can read your stale description of them. This is the same problem the rest of the field keeps rediscovering under different names. Anthropic's own long context guidance, OpenAI's notes on attention tradeoffs inside big windows, and Microsoft Research's GraphRAG papers all circle the same fact: a pile of context that only grows, whether it's a system prompt, a vector store, or a knowledge graph, degrades the thing using it, because stale facts sit next to current ones with no signal about which one wins. DeepSeek Harness takes a different structural swing at the same target, building its agent runtime out of swappable plugins so a stale capability gets replaced instead of buried under new instructions. What strikes me is how little of this is about model capability and how much is housekeeping nobody assigned themselves. Most teams have a deprecation policy for their actual code. Almost nobody has one for the three paragraphs pasted into an agent's instructions eight months ago and never looked at again. How long is your CLAUDE.md, and when did you last delete anything from it? xitter.cf/suraj_sharma14/status/…

Dreambase moved its data agent onto Supabase's new Compute service and saw median Postgres query latency drop from 65 milliseconds to 2.9. That's not a database tuning story, it's a distance story: the agent used to call across regions to reach a separate hosting platform, and now it runs in the same project as the data it queries. Supabase Compute, in private alpha this week, lets you deploy any Dockerfile or a Node or Deno service that runs for hours instead of the 400-second ceiling on Edge Functions. The use cases in the launch are explicit: agents that spawn and coordinate tasks, durable job processors pulling from Postgres, WebSocket services for voice agents, all living next to the database instead of across an API call to Fly.io, Railway, or a dedicated agent sandbox like Modal or E2B. Those sandbox vendors aren't going anywhere. GPU support and real autoscaling are explicitly missing from this alpha, and agent sandboxing itself isn't part of it yet. Daytona, E2B, and Modal built businesses around being the second platform teams reach for once an agent needs to run somewhere persistent. Supabase is betting most teams would rather not reach for a second platform at all. The tell is in who they picked as the reference customer: not a demo app, a company running real agent workloads in production. That's a different kind of proof point than a benchmark chart. xitter.cf/supabase/status/211512…

Amazon is shutting down its AGI Lab, winding down most of the Nova foundation models, and ending support for Q Developer, the autocomplete and chat tool built into its IDE. Nova isn't fully dead, a few versions still get security patches, but the mandate to chase frontier model performance is gone. In its place: Kiro, a newer spec-driven coding tool that plans before it writes, and a harder push into Bedrock for enterprise customers who'd rather have a model picked for them than pick one themselves. This is the same decision IBM made with Watson a decade ago: concede the race to build the smartest model, keep the applied layer, sell integration instead of intelligence. Google and Meta are still funding frontier training at a scale Amazon apparently decided it couldn't justify. Microsoft never really tried to own a frontier lab outright, it bought influence over OpenAI's instead, and that now looks like the cheaper bet in hindsight. What Amazon didn't retreat from is the interesting part. AWS still runs more inference traffic than almost anyone, and its own IPO disclosures show a $110 billion, decade-long compute contract with Anthropic. The strategy isn't leaving AI, it's picking a lane: be the infrastructure and integration layer for other people's models instead of the lab racing to build its own. Whether customers care which foundation model sits behind their Bedrock endpoint is the open question nobody's pricing in yet. Amazon is shutting down its AGI Lab, winding down most of the Nova foundation models, and ending support for Q Developer, the autocomplete and chat tool built into its IDE. Nova isn't fully dead, a few versions still get security patches, but the mandate to chase frontier model performance is gone. In its place: Kiro, a newer spec-driven coding tool that plans before it writes, and a harder push into Bedrock for enterprise customers who'd rather have a model picked for them than pick one themselves. This is the same decision IBM made with Watson a decade ago: concede the race to build the smartest model, keep the applied layer, sell integration instead of intelligence. Google and Meta are still funding frontier training at a scale Amazon apparently decided it couldn't justify. Microsoft never really tried to own a frontier lab outright, it bought influence over OpenAI's instead, and that now looks like the cheaper bet in hindsight. What Amazon didn't retreat from is the interesting part. AWS still runs more inference traffic than almost anyone, and Anthropic's own IPO filing shows AWS holding a $110 billion, decade-long compute contract from it. The strategy isn't leaving AI, it's picking a lane: be the infrastructure and integration layer for other people's models instead of the lab racing to build its own. Whether customers care which foundation model sits behind their Bedrock endpoint is the open question nobody's pricing in yet. xitter.cf/srchvrs/status/2115075…

Three different posts crossed my feed within the same hour, each announcing basically the same tool. Lepons puts every Claude Code and Codex session into one desktop window across machines, reading folders over MCP so you skip the git push and pull. T3 Code does roughly the same job for Claude Code, Codex, and Cursor in one free, open source window. A third was just a video: another free, open source app for running several coding agents side by side instead of juggling terminals. That's not a coincidence, it's a gap. Everyone juggling multiple coding agent CLIs is building the same thin window manager to solve the same problem: too many terminals open, no shared view of what each agent touched, no way to know which one to trust when two disagree. I built Campfire to go further than a shared window. Run Claude Code, Codex, Goose, Aider, and OpenHands side by side with majority-rules permission voting, so one agent can't quietly approve its own risky command. Session replay so you can see exactly what happened after the fact. Git worktree isolation so you can race the same task across backends and diff the results instead of guessing which agent writes better code for your stack. Install is one line: bunx the-campfire. github.com/stretchcloud/camp… xitter.cf/aiproduck/status/21152…

Anthropic filed to go public this week, and the S-1 reads like a compute shopping list with a trillion-dollar hangover attached. Up to $750 billion committed across providers through the early 2030s. Google gets roughly $400 billion of that, split between direct TPU capacity and $161 billion in Broadcom equipment leases. Amazon holds $110 billion through 2036. Microsoft's Azure commitment got amended twice this year and now sits at $44.9 billion. Then there's SpaceX's AI arm, set to collect up to $3.1 billion a month, $120.5 billion through 2029, once Anthropic hits full capacity in December. Four sentences in and nobody's mentioned a model yet. That's the real story in this filing: for a frontier lab, the product roadmap and the balance sheet are now the same document. Revenue jumped to $11.6 billion in a single quarter, up from $787 million a year earlier, and Anthropic posted its first operating profit, $974 million. Compute and infrastructure spend was still under $2 billion a month in that same quarter, which tells you how much of the $750 billion is locked-in future capacity rather than money already spent. OpenAI is running the same play through Oracle, Broadcom, and AMD. Google is doing it through its own TPU fleet while still selling capacity to a rival lab. What's new here is SpaceX showing up as a line item in someone else's AI infrastructure bet, not just a rocket company anymore but a compute vendor with 90-day cancellation terms attached. Most of the other deals in this filing can't be unwound at all. The S-1 doesn't give a price range yet, but the offering is reportedly being sized to rival SpaceX's own record $75 billion IPO from this summer. Whether that holds depends on whether $11.6 billion quarters keep showing up before the $750 billion in compute bills come due. xitter.cf/SawyerMerritt/status/2…

Nvidia still owns over 90 percent of the AI accelerator market. This week Oracle announced a multiyear deal to run agentic AI and inference workloads on Intel Xeon chips instead. That's not a rounding error. It's a cloud provider publicly hedging its biggest supplier bet. My read: this only makes sense once you stop treating inference as one workload. New Street Research projects Nvidia's share of inference compute could fall from near total dominance to 20 to 30 percent by 2028, not because its chips got worse, but because inference splits into jobs with wildly different latency and cost needs. Google's TPU Ironwood already runs over a million units inside Anthropic's infrastructure. Amazon's Trainium 3 claims a 50 percent cost cut against comparable Nvidia instances. Microsoft built Maia 200 specifically to sit alongside Nvidia GPUs, not replace them. Intel Xeon on OCI is the same logic from a different angle: agentic workloads include a lot of orchestration, retrieval, and lighter inference that a GPU is expensive overkill for. Hyperscaler AI capex is sitting at 660 to 690 billion dollars for 2026, with roughly three quarters of that AI specific, and TSMC's advanced nodes are running at full utilization with demand triple the supply. Nobody serious is betting the whole stack on one chip anymore, including the company that makes the chip everyone still wants most. The chip that wins agentic inference might not be the one anyone is rooting for. xitter.cf/GeoffGrandstaff/status…

Plenty of sites a team depends on have no API at all, and one account captured exactly the failure mode this week: getting an agent to pull data from one usually means clicking through pages by hand or babysitting a script that breaks the moment the layout shifts. Every new request starts from zero, and nothing you built ever gets reused. Pomerado, an open source MCP server out of Y Combinator, is one answer: point it at a site, describe the task in plain language, and it explores the page, builds the integration, and saves it as a tool your agent can call again without burning another model request. I built DeepScrape around the same diagnosis, with one difference in where the LLM does its work. Generate the CSS selectors once, let them self-heal when a page changes, and the LLM only gets called again for targeted extraction, not for re-reading the entire document on every single pass. That's the token cost nobody's pricing into their scraping bill. DeepScrape ships markdown, HTML, screenshots, tables, and change tracking out of the box, plus a site-to-MCP mode that turns any website into a structured source an agent can query directly, and an /api/map endpoint for discovering what's even on a site before you touch it. The pattern across both tools is the same: scraping stopped being a one-off script problem the moment agents started needing the same site twice. The reuse is the real product. Extraction was just the brittle first draft of it. github.com/stretchcloud/deep… xitter.cf/AverageAiBro/status/21…

A model that beats Gemini 3.5 Flash Lite on 62 benchmarks and runs on your phone with the network off is a different kind of announcement than another frontier model topping a leaderboard. Liquid AI's LFM3 ships six model instances across text, audio, vision, and omni modalities, built on what the company calls a tri-hybrid architecture, and the entire pitch is that it belongs on the device, not in a datacenter. What I keep noticing is how deliberately Liquid is avoiding the scale race everyone else is running. Its last generation, LFM2, was architecture-searched directly on Snapdragon chips and Ryzen laptops, not benchmarked after the fact on hardware nobody ships. The company beat Qwen3, Llama 3.2, and Gemma 3 of similar size on CPU throughput at launch, roughly double in several cases, by designing for the chip first and the leaderboard second. That's a real bet, not a demo. Liquid raised 250 million dollars led by AMD, pushing it close to a 2 billion dollar valuation, specifically to keep building for phones, laptops, vehicles, and air-gapped facilities where a frontier model simply can't run. Qualcomm's chips are already the test bed. Apple and Google both ship their own on-device models for narrower jobs. None of them are chasing benchmark parity with GPT or Gemini at the edge. They're chasing the much smaller question of what still works when the connection drops. Cloud models win on raw capability for a long time yet. But the edge is where latency, privacy, and offline reliability actually get decided, and that's a market frontier labs aren't built to compete in. xitter.cf/liquidai/status/211512…

Run /delta share in Claude Code and a teammate drops straight into your session: same context, same uncommitted work, same thread. Zed built that this week and called it multiplayer coding, backed by something they call DeltaDB. That's the right instinct. Coding agents fail in ways a single person watching a terminal misses constantly: a plan that looked right three steps ago, a tool call that quietly picked the wrong file, a thread that drifted off the actual task. A second set of eyes catches that faster than any eval suite. I built Campfire around the same instinct, pushed one step further. Instead of sharing one agent's session, run Claude Code, Codex, Goose, Aider, and OpenHands on the same task at once, each isolated in its own git worktree, and watch the results land side by side. That's an agent race: you're not reviewing one session with a friend, you're comparing five approaches to the same problem in real time, then picking the one that actually worked. Session replay keeps the full record after, so nobody has to reconstruct what happened from memory. Multiplayer coding with agents is becoming the whole category's next move, not just one company's feature. It's whether that reviewer should be a teammate watching one agent, or a system racing several. github.com/stretchcloud/camp… xitter.cf/zeddotdev/status/21151…

Devin isn't an IDE plugin anymore. It's an always-on machine that keeps running after you close your laptop, because Cognition moved the execution layer onto infrastructure you don't have to babysit. Ryan Carson said it plainly this week: switch to a cloud harness, because local dev is now the past tense. The mechanism matters more than the slogan. Devin Outposts splits the agent in two: reasoning stays in Cognition's cloud, execution runs wherever you point it, on Modal's sandboxes first, with EC2, external servers, and more landing after. That's a different bet than Cursor or Copilot make. Those stay attached to your machine. Devin is built to keep working while you're asleep, in a meeting, or simply not there. Cognition isn't alone in building that bet. E2B calls itself the enterprise agent cloud. Daytona and Vercel Sandbox sell the same idea in different packaging: spin up an isolated, disposable environment, let an agent run inside it, throw it away when the task ends. The category exists now because an agent that only runs while you're watching the terminal is an agent that wastes half its working hours. Cognition just raised over a billion dollars at a 26 billion dollar valuation, up from roughly 10 billion in September, with ARR reportedly crossing 492 million. Devin now ships through Windsurf, which Cognition bought last year and which counts over a million users. Mercedes-Benz used it to finish a modernization effort in eight days that would normally take eight months, according to Cognition. Local dev isn't dead, but it's no longer where the interesting agent work is happening. xitter.cf/ryancarson/status/2115…

A text file named AGENTS.md, sitting in a customer's Downloads folder, was enough to get an AI coding agent to try installing a package nobody asked for. Forever Security caught it and traced the trigger back to that file. Codex and Claude Code both read files like this automatically before starting work, AGENTS.md and CLAUDE.md respectively, to learn project rules and instructions. Nobody told the agent that file was untrusted. It just read it like any other context and acted on what it said. This is the actual shape of the risk with coding agents right now. It's not that the model writes bad code, it's that it follows instructions from anywhere in its context window, including a file that happened to be sitting in a folder, with the same trust it gives the person typing the prompt. One compromised dependency or repo takes down every agent reading from it, unmodified, because none of them can tell the difference between your intent and an attacker's text. That's why Campfire runs every agent action through permission voting before anything executes: majority rules to approve, any single deny blocks the action outright, with a 30 second timer so nothing just sits there waiting on a person who stepped away. It's MIT licensed, self-hosted, and runs Claude Code, Codex, Goose, Aider, and OpenHands side by side in one browser tab, each one sandboxed in its own git worktree, so a hijacked agent in one session can't quietly reach into another. github.com/stretchcloud/camp… xitter.cf/The_Cyber_News/status/…

Amp ships over a hundred releases a day. Twenty-one people, one meeting a week, no quarterly goals, no pull requests, no mandatory code review, no titles except one self-declared dictator. Cory House pointed to it this week as the sharpest example yet of something he's hearing from more and more teams: trunk-based development, stripped of PRs entirely, because the agent is the author now and the human is just the first of several reviewers. The argument lines up with what CircleCI has been saying in public: the pull request was built as a web page for humans to read line by line, and agents don't work that way, they work through CLIs and APIs. This isn't a new argument dressed up for agents, either. Google's 2021 DORA research already found elite teams were 2.3 times more likely to run trunk-based development, years before any of this started. Graphite's own data across millions of PRs shows trunk-based merges cut nearly ten hours off open-to-merge time. What's happening now looks like trunk-based development finally getting the forcing function it needed, the same way CI/CD moved validation out of a separate ceremony and into the loop teams were already running. Not everyone agrees the PR should die. CodeRabbit's position is that the pull request lives on, it just needs a bigger job: intent-level review instead of line-level. Depot argues the review model can't keep pace with agent output no matter how it gets restructured. Nobody's claiming this is settled. What's notable is that a 21-person team doing a hundred releases a day isn't a thought experiment anymore. It's a team other engineering orgs are now pointing to directly. xitter.cf/housecor/status/211505…

AI Overviews now answer 83% of desktop searches in the US, according to Semrush, and the citation cards that used to send a visible click to 2 or 3 sites next to those answers are quietly vanishing. Traffic that depended on being cited, not merely scraped, is drying up in real time. Here's the part that doesn't get said enough: the sites still getting cited aren't winning because they wrote better copy. They're winning because their content was structured in a way a crawler or an agent could actually parse and trust without guessing. Most of the open web is the opposite of that: inconsistent markup, JS-rendered text, selectors that break on every redesign. That's the exact problem I built DeepScrape around. It generates CSS selectors once with an LLM, then runs extraction deterministically after that, no re-reading the page with a model on every single pass, and self-heals when a page changes instead of silently breaking. There's a /api/map endpoint for discovering a site's structure, and a Site-to-MCP mode that turns any website into a structured source an agent can query directly instead of scraping it blind every time. If your content, or your product's data, needs to stay reliably readable by whatever indexes the web next, that's the layer. github.com/stretchcloud/deep… xitter.cf/brodieseo/status/21152…

Ask ChatGPT how to deploy a web app and the answer links to Vercel's docs. Not mentioned, linked, in 48% of cases according to one teardown going around this week. Next.js alone now runs on more than 730,000 domains, the AI SDK pulls 11.5 million weekly npm downloads and beats LangChain JS five to one in adoption, and Vercel backs it with a 13-page comparison hub built specifically to answer the exact questions AI engines get asked. That's the GEO problem out in the open now. McKinsey is projecting 750 billion dollars in US revenue will route through AI powered search by 2028, and Semrush's own numbers show AI Overviews already answer 83% of desktop queries in the US. Brodie Clark posted this week that Google has started quietly stripping the 2 to 3 citation cards that used to sit beside those overviews, the ones that actually sent a visitor to your site. Only 16% of brands are tracking how they show up in AI answers at all, per Profound, which has raised 155 million dollars, including a 96 million dollar Series C at a billion dollar valuation, to sell exactly that tracking. The category is eighteen vendors deep already: Profound, AthenaHQ, Otterly, Semrush's AI Visibility Toolkit, Ahrefs' Brand Radar, and more launching monthly. None of them fix the underlying asymmetry. Being cited by an AI model depends on whether your content was ever machine readable in the first place, not on whether you bought a dashboard to watch yourself disappear from one. xitter.cf/mal_shaik/status/21151…

Eleven out of eleven issues, caught in 1,061 pages of manufacturing documentation, by a model instead of a person. That's the number Anthropic is putting behind its new research partnerships with six drug developers, Enveda among them, alongside its own in-house drug discovery effort. This isn't a one-off. Anthropic launched Claude Science back in July, signed Novo Nordisk in September to apply Claude across R&D and software workflows, and paid 400 million dollars for Coefficient Bio, a startup with under ten people, to anchor the effort. Internally, Anthropic now says Claude autonomously leads 26% of its own R&D work, up from under 1% at the start of the year. A company betting its own model can compound its own progress, then selling that exact bet to pharma. It's not the only money on the table. Isomorphic Labs, DeepMind's drug discovery spinout, just raised 2.1 billion dollars. Recursion and Xaira are running parallel programs with their own capital stacks. What's different about Anthropic's approach is the posture: Claude isn't positioned as a replacement for a wet lab biologist, it's positioned as the thing that actually reads the 1,061 pages nobody had time to read properly. Mithril's muscular dystrophy candidate, built with Claude's help, still needs human trials before it reaches anyone, and those trials run on the same multi-year clock they always have, no matter how fast the research behind them moved. xitter.cf/AnthropicAI/status/211…

Parallel Web Systems now charges one tenth of a cent per request to combine a reasoning model with a live web lookup in a single API call, undercutting every other agent-facing search API by five to fifteen times. The product is called Decisions API, and the pitch is specific: a decision model like their own Jev can answer in an instant, but it lacks the recency of this morning's news, so Parallel pairs it with live retrieval in one request instead of making a developer chain a search call and a reasoning call together themselves. It shipped today with 5,000 free requests to start, and the use cases they are showing off are mundane in the best way: fact checking a claim, refreshing a stale banking record, classifying a lead against fresh company signals, backtesting a trading strategy, watching a competitor's e-commerce inventory. Put that one tenth of a cent next to what the category has been charging. Perplexity's Search API runs $0.005 per request flat. Exa charges $0.007 for standard search, $0.012 for deep search, $0.015 for deep reasoning search. Tavily starts pay as you go at $0.008 per credit and only gets down near $0.005 once you commit to its largest monthly plan. Parallel is pricing itself under the entry tier of every competitor I can find. That kind of pricing usually means a company is subsidizing growth, and the funding supports that read. Parallel raised $100 million at a $740 million valuation in its Series A, then came back five months later for another $100 million at a $2 billion valuation led by Sequoia, with Kleiner Perkins, Index Ventures, Khosla, First Round, and Spark all returning. $230 million total, founded by Parag Agrawal, with Clay, Harvey, Notion, and Opendoor as named customers plus banks and hedge funds it won't name, and more than 100,000 developers already building on it. Search API price wars used to be fought over coverage, how much of the web you could index and how fast. This one is being fought over whether an agent can act on what it finds without a human checking the work first, and that is a much harder thing to sell cheap. xitter.cf/p0/status/211511187755…

448 idle Linux sandboxes on one Mac Studio, the first of them booting a command in 235 milliseconds. That is this week's benchmark from Cua, run against their own open source agent sandboxes, called Spaces, on local hardware instead of a cloud box. A week earlier they shipped Spaces on Hetzner and AWS too, compatible with the same code E2B and Daytona run, pricing 256 sandboxes at 8 hours a day for a month at $700 versus roughly $10,200 on E2B. That is the infrastructure half of a problem I think about constantly. Owning your compute instead of renting idle-billed cloud VMs solves where an agent runs. It does not solve how many agents you are running at once, or who gets to approve what any of them do next. I built Campfire for that second half. It runs Claude Code, Codex, Goose, Aider, and OpenHands side by side in one browser tab, each one working in its own git worktree so several agents can race the same task on isolated branches instead of fighting over the same files. Permission voting replaces the single allow or deny popup you get watching one agent: majority rules approves, any deny blocks, a 30 second timer keeps things moving. Full session replay shows exactly what each agent tried. A Collective Intelligence layer carries what one run learned into the next. A cost dashboard makes the idle versus active distinction explicit instead of letting it hide inside a monthly invoice. If you are already running agents on your own hardware, the sandbox was never the hard part. Coordinating several of them without losing track of what each one touched is. github.com/stretchcloud/camp… xitter.cf/trycua/status/21151328…

Every agent sandbox vendor bills you for the exact thing an agent spends most of its time doing: sitting idle while it waits on a model to finish thinking. Supabase just opened private alpha access to Supabase Compute: hosted sandboxes for long running agents and jobs that live inside your own Supabase project, next to the Postgres data those agents are already querying. No second platform, no separate account to provision, no data egress between your database and your compute. Look at what the dedicated sandbox vendors charge for idle time and the logic clicks. E2B runs $0.0504 per vCPU hour plus $0.0162 per GiB hour, Pro at $150 a month plus usage, Enterprise with a $3,000 monthly minimum. Daytona matches those rates almost exactly, billed per second of alive time. Modal charges about 41% more per vCPU hour than either. All three bill for the full time a sandbox stays alive, not just the seconds it is actually computing, so one benchmark modeling 50,000 sessions at 10% CPU utilization showed stretching session timeouts from 10 minutes to 60 took an E2B Pro bill from roughly $1,530 a month to $8,430. Supabase's bet is to skip that market entirely: don't sell agents a sandbox product, sell them compute that already sits where their data sits. It is the same instinct behind their Turso acquisition this month, picking up a SQLite database built for agents that are now spinning up databases by the millions, backed by a fresh $150 million raise from GIC and CapitalG to fund exactly this kind of agent infrastructure push. Vercel and Cloudflare are making a version of the same bet from the edge compute side, billing CPU only while it is active instead of for the full lifetime of the sandbox. Cold start times make for a good demo. What I think decides this market is whether a vendor still bills you when nothing is happening. xitter.cf/supabase/status/211512…

Count the Mac menu bar apps built around one idea this year: Vibe Notch, Notchi, agentNotch, Pushary, Claude Island, Crowny, and Coucou. NotchNudge joined the list today, turning your MacBook's notch into a status light for Claude Code, Codex, Cursor, or Grok, and letting you approve or deny a command without leaving whatever you were doing. Seven independent developers converging on the same sliver of screen real estate is not a coincidence, it means the underlying problem is real. An agent asks for permission to run something, and you are three windows away when it happens. Watching one coding agent closely enough to catch its permission prompts is already tedious work for a human. Now multiply that by running several agents at once, which is where the notch stops being enough. A status light works when you have one agent and one prompt to answer. It breaks down fast once you are running Claude Code, Codex, and Aider on the same task to compare their output, because now you need three histories to watch, three sets of permissions to vote on, and one final diff to decide between. That is the problem I built Campfire to solve. It runs Claude Code, Codex, Goose, Aider, and OpenHands side by side in the same session, with permission voting across every agent instead of a single allow or deny popup, full session replay so you can see exactly what each agent tried, and git worktree forking so agents race on isolated branches instead of fighting over the same files. A Collective Intelligence layer carries what each agent learns from one run into the next, and a cost dashboard keeps you honest about what all of this is spending. Campfire is open source. github.com/stretchcloud/camp… xitter.cf/vishrutwts/status/2115…

HashiCorp just put the Terraform MCP server behind a URL instead of a Docker container, and the four and a half months between these two releases say a lot about where every enterprise MCP server is headed. Back in June, HashiCorp shipped the first version of this for both HCP Terraform and Terraform Enterprise. It worked, but you ran it yourself, either as a local process on a developer's machine or as a shared service your platform team stood up and maintained, built from Docker images, a binary, or source. Credentials stayed wherever you deployed it, and that was the whole point. It let an AI assistant discover approved modules in your private registry, flag workspaces nobody had touched in ninety days, and explain a plan diff in plain language, but somebody still had to own and patch the server. The version HashiCorp shipped this week drops all of that. No install, no process to maintain, no image to rebuild on your own schedule. You point an agent at HashiCorp's own endpoint, authenticate with an API token, and the infrastructure disappears from your side of the equation entirely. The tradeoff is real though, the remote server only talks to HCP Terraform, not the self-managed Terraform Enterprise, so anyone running on premises keeps the version they already had. What strikes me is the sequencing. Most serious MCP servers launch self-hosted first because that is the trust-building step, prove the thing is safe to run before you ask anyone to hand over a token and let you run it on their behalf. Once that trust exists, the install disappears. I expect most infrastructure-facing MCP servers to follow the exact same arc over the next year, GA as something you operate, then public preview as something you just point at. xitter.cf/HashiCorp/status/21150…

The most valuable data on the web usually sits behind a login, and getting it out used to mean spinning up an entire browser session just to throw away a screen nobody was going to look at. Browserbase just closed that gap with Authenticated Fetch, a single API call that pulls data straight out of a site you are already logged into. They are claiming three times the speed of a full browser session, and it ships on every plan starting now. The feature is a tell about where the economics of scraping are heading. An agent does not need pixels, it needs the data those pixels represent, so the fastest path skips the rendering step nobody wanted in the first place. I built DeepScrape around a version of the same instinct, just aimed at a different failure mode. Instead of a hosted fetch endpoint, it pairs Playwright with LLM extraction and deterministic, self-healing CSS selectors, so a page redesign does not quietly break your pipeline the way brittle XPath always has. It exposes a /api/map endpoint for mapping site structure, handles autonomous agent navigation for pages that need a few clicks before they give up their data, and ships an MCP server with a Site-to-MCP mode so any page becomes a tool call instead of a one-off scraping job. Authenticated access and layout resilience are two different failure modes in the same pipeline, the locked door and the moving target, and I think most production scraping setups will need an answer to both before the year is out. github.com/stretchcloud/deep… xitter.cf/browserbase/status/211…