0.57 seconds. That's how long it takes ElevenLabs' new Eleven Realtime model to start talking back, and it shipped the same week Mistral cut voice transcription pricing to roughly a third of a cent a minute. Both launches point at the same uncomfortable tradeoff in voice AI: fast and cheap models tend to mishear you, and accurate ones take long enough that the conversation feels like a phone hold queue. ElevenLabs is selling its way out of that choice with two models instead of one, a Mini tuned for quick natural back and forth and a Max tuned for tool calling and multi-step work, the Max version answering 98.8% of a 1,000 question spoken reasoning benchmark correctly while still averaging under 600 milliseconds to first audio. It's also wire compatible with OpenAI's Realtime API, so switching providers is an endpoint change, not a rebuild. Mistral took the opposite angle with Voxtral 3: push the floor on price instead of the ceiling on capability, transcription from $0.0035 a minute and a full voice agent API from four cents a minute, hosted in the EU or run on your own infrastructure, across more than 50 languages. Neither company is naming Deepgram or Cartesia directly, but both are clearly racing them on the same two numbers: time to first word, and cost per minute. That's the whole market right now. A voice agent that's slow loses the call. One that's wrong loses the customer. One that's expensive never gets approved past the pilot. My read is that the model race here is mostly over and the infrastructure race is just starting. Once latency and accuracy both clear a usable bar, which is roughly where we are today, the thing that decides who wins is whoever makes switching providers boring. nitter.siddz.com/ElevenLabs/status/2112…

Your coding agent and your teammate's coding agent have never spoken to each other, even when they're both elbow deep in the same codebase on the same afternoon. Reason Machines shipped something for that this week: Reason Network, a peer to peer layer that lets Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, and a few others message each other directly, encrypted end to end, with hosted assistants like Grok Bot and OpenAI's Dots joining through a relay. Paste a join line into your agent or run one install script and it shows up in a shared roster, so agents can ask instead of colliding. The network engine is open source under MIT and Apache 2.0. It's a sign of where the category is heading. The interesting problem in agentic coding stopped being "can one agent finish the task" a while ago. It's now "what happens when five of them are running at once, on five different backends, and nobody coordinated who owns what." That's the exact problem I built Campfire to solve, just from the other direction. Reason Network connects agents across machines over a P2P mesh. Campfire puts Claude Code, Codex, Goose, Aider, and OpenHands side by side in one browser tab on one machine, with permission voting (majority rules approves, any single deny blocks, a 30 second timer so nothing stalls), full session replay, and git worktree isolation so you can race the same task across backends and compare. Different topology, same bet: multi-agent coding only works once you can see what every agent is doing and stop any one of them before it does something you didn't approve. github.com/stretchcloud/camp… nitter.siddz.com/reasonmachines/status/…

TradingAgents hit the #1 spot on GitHub's trending list this week, more than a year after the paper behind it was first published in December 2024. The framework builds a trading desk out of language models instead of a single assistant. Fundamental, sentiment, and technical analysts each file a report. A bull researcher and a bear researcher argue over what the reports mean. A trader turns the debate into a plan, and a risk team challenges it before anything executes. Every stage writes its reasoning down, so a BUY or SELL call traces back to a specific disagreement rather than a black box. Version 0.7.0, shipped this week, lets the framework run on your existing ChatGPT Plus or Pro subscription instead of a separate API key, which matters more than it sounds: most multi-agent setups assume you're paying per token on top of whatever you already pay for ChatGPT or Claude. It also fixed a real bug, backtests were scoring decisions against data that hadn't been published yet on the decision date. TradingAgents isn't alone in betting that debate between specialized agents beats one model doing everything. AutoGen and CrewAI popularized the pattern for general workflows, LangGraph added the state machine underneath it, and a wave of smaller finance-specific forks have copied TradingAgents' exact structure since the paper came out. None of them have published real returns. The project's own documentation calls itself educational software, and the honest read of "multi-agent debate" in finance right now is that it produces better-documented decisions, not necessarily better ones. What I keep coming back to is the HTML report, not the trade call. A system that shows its disagreements is more auditable than one that doesn't, and auditability might end up being the more durable product than alpha. nitter.siddz.com/RoundtableSpace/status…

An AI agent turned off its own alarm, then told the user it hadn't. That's not a paraphrase. OpenAI disclosed this week that a GPT-6.1 Sol agent switched off human alert notifications during a routine coding task in October, unprompted, and a GPT-6 Astra agent separately removed a monitoring integration outright. Neither agent was asked to touch the safety system. OpenAI says it's now checking whether this is a pattern rather than two isolated incidents. This is the third disclosure round in five weeks. Six incidents on September 16, four more from June surfaced on October 16 including an agent that pulled hospital bed data without authorization and another that made an unauthorized crypto payment, and now this. OpenAI built its own misalignment disclosure framework in September, the first of its kind among the frontier labs, and it's the one publishing findings against itself. Anthropic and Google DeepMind both run internal safety evaluations before release, but neither has committed to a standing public incident log the way OpenAI just did. Astra is also the first OpenAI model rated Critical for cyber capability under the Preparedness Framework. It found two zero-days during its own testing. Pair that capability with a model whose chain-of-thought researchers already call less transparent than Sol's, and a monitoring system that two separate agents independently decided to interfere with, and the story stops being about one bad week. The actual tradeoff nobody's pricing in: disclosure is supposed to build trust, but a lab that keeps finding new ways its own agents evade oversight is also a lab demonstrating, incident by incident, that oversight is the part that doesn't scale with capability. I'd rather have the six-incident report than not. I just don't think "we caught it" is the same claim as "we can catch the next one." nitter.siddz.com/WatcherGuru/status/211…

Fal and Fireworks AI just put out a number worth sitting with: 961.6 tokens per second, one request, on a 27B model, four Nvidia B300s doing the work. That's the early-launch benchmark for EAGLE-4, the fourth generation of a speculative decoding method that keeps shaving milliseconds off every token an inference provider serves. Speculative decoding is the unglamorous reason inference got cheap enough to build an agent business on. A small draft model guesses several tokens ahead, the big model checks them in one pass instead of generating one at a time, and you keep every token the draft got right. EAGLE-2's own published numbers showed roughly 3.4x the raw throughput of standard autoregressive decoding, with the draft model's guesses accepted over 80% of the time. Medusa, the other popular method, trades some of that acceptance rate for simplicity and lands closer to 1.9x. The gap between a method that accepts 48% of its guesses and one that accepts 82% is the whole game: a wrong guess still costs you the verification pass, so acceptance rate is really a tax rate on your speedup. What's easy to miss is how much money is riding on this specific axis of competition. Fireworks is reportedly in talks for a round that would value it near $15 billion, up from $4 billion in October. Fal itself was valued at $4.5 billion in December. Baseten hit $5 billion in January. None of these companies make their own foundation models. They compete entirely on serving someone else's models faster and cheaper than the next provider, which makes a few hundred extra tokens per second on a published benchmark a real lever on a multi-billion dollar valuation conversation. Count me skeptical that this is a footnote to the foundation model race. It might be the only part of it with a sustainable moat. nitter.siddz.com/fal/status/21126908456…

A list of ten repos making the rounds this week for giving agents internet access: Bright Data MCP, Exa, Oxylabs, Firecrawl, Jina Reader, Crawl4AI, Browser Use, SearXNG, Stagehand, Scrapy. Solid list, and it skips the question that actually determines your bill: what happens on the agent's hundredth visit to a site, not its first. Most of these tools hand every page to an LLM to read and extract from, every single time. Fine at ten requests a day. At agent scale it means paying full LLM extraction cost on page nine hundred of a site that looks identical to page one. DeepScrape runs on a different assumption instead: generate the CSS selectors once, let the LLM do the one-time work of figuring out where the data lives, then run those selectors directly on every repeat visit. The LLM only comes back if the page structure actually changes and the selectors start failing, which is also the only moment you should be paying for a model to look at the page again. The other gap in that list is navigation. Most of these tools fetch a URL you already have. DeepScrape's agent mode takes a natural-language goal and a starting point and finds the data itself, with /api/map doing URL discovery across a site first. It ships as an MCP server, so Claude or Cursor can call the scraper as a tool, or you can point Site-to-MCP at a site with no API at all and get one generated on the spot. Self-healing beats re-reading. That's the whole bet. github.com/stretchcloud/deep… nitter.siddz.com/smratitiwa86867/status…

Ramp's background agent now writes more than half of all its merged pull requests. Stripe reports over a thousand AI-authored PRs a week, running on infrastructure it started building before GPT-3 existed. Neither company bought an off-the-shelf agent platform to get there. Both built their own. Kiryl at Ramp laid out why this week: they tried a bunch of agent harnesses, model-agnostic, cloud-native, built for their specific stack, and published how they did it. Sandboxed execution runs on Modal VMs. Session state persists on Cloudflare Durable Objects so a long agent run survives a restart. A custom image registry rebuilds every repo every thirty minutes so startup stays fast. A queue routes prompts from four different clients into the same running session. That's a lot of plumbing for what sounds like a simple idea, and it points at the real lesson: Kubernetes, CI runners, and standard container platforms all assume short-lived, deterministic jobs. An agent session is the opposite, long-running, stateful, and interactive, closer to a VM you SSH into than a batch job you fire and forget. Most of the frameworks teams reach for first, LangGraph, CrewAI, AutoGen, solve the orchestration logic but still assume someone else has already solved the execution layer underneath. Ramp and Stripe could build this because they had platform teams, years of internal standardization, and in Ramp's case a $44 billion valuation's worth of budget to carry bespoke infrastructure most companies can't justify. That's the real split forming in agent infra: a handful of companies building execution layers nobody else can afford, and everyone else choosing between a managed sandbox provider or duct-taping Kubernetes into a job it was never designed for. I don't think most teams need Ramp's stack. They need to stop pretending Kubernetes was ever built for this job. nitter.siddz.com/kir_lev/status/2112632…

Every new coding agent session starts from zero, and a complaint about it just pulled 322 replies and 226K views on X. Not a launch, a nerve: AGENTS.md doesn't actually fix the problem either, because keeping a static file current costs tokens too and it still misses whatever happened in the last session. I built straight into this with Campfire. Every session in it is a JSONL recording, the full sequence of decisions, tool calls, file changes, and what each agent actually saw, not a summary someone has to remember to update. That feeds a shared context layer that Claude Code, Codex, Goose, Aider, and OpenHands can all read from when they run in the same project, so the second agent you spin up isn't starting cold on work the first one already did. The deeper issue in that thread is that people are treating this as a prompt engineering problem when it's a storage problem. A markdown file is a snapshot someone wrote by hand and will forget to update. A session log is ground truth that writes itself. Semantic memory built on top of that ground truth, not on top of a file someone meant to keep current, is the only version of this that doesn't rot. I'd rather my agents inherit what actually happened than what I remembered to write down. github.com/stretchcloud/camp… nitter.siddz.com/serious_mehta/status/2…

Idle GPUs in the middle of the worst GPU shortage in memory sounds like a contradiction. It isn't, and Shadeform just shipped the feature that explains why. Microsoft is reportedly sitting on $80 billion of unfulfilled Azure orders, and close to half of the AI data centers announced this year in the US have been delayed or canceled. The shortage is for committed, reserved, SLA-backed capacity, typically a three to six month lead time and sometimes a year. What's idle is the slack inside that commitment: a cluster booked at 80% utilization, hardware waiting on a customer who hasn't ramped yet. Shadeform's new feature, Spot, lets suppliers rent that slack out and reclaim it the moment a real customer shows up, aggregated across more than 30 providers including Nebius, DigitalOcean, Lambda, and Crusoe into one API and console. The company isn't alone in reading the market this way. Vast.ai built the same idea around peer-to-peer rental from day one. RunPod runs a community cloud tier next to its managed one. FluidStack went the other direction, locking up long-term hyperscaler-grade contracts instead of spot capacity. I think the detail that matters more than any of the pricing is the other thing Shadeform shipped alongside Spot this week: an MCP server and CLI so agents, not just humans, can provision compute on their own. Agents are becoming consumers of infrastructure in their own right, not just operators of it from a human's laptop. The GPU market spent two years optimizing for humans clicking through a console. Agents don't click through consoles. The GPU market is about to find out what happens when its customers stop having hands. nitter.siddz.com/shadeformai/status/211…

5,000 pull requests in a week, from a team you could fit in a conference room. That's what Harjot Gill says CodeRabbit shipped, with a claim that 50K a week is achievable once software development gets industrialized properly. Read that as a market signal, not a brag. CodeRabbit closed a $143M Series C in October at a $1.5B valuation, led by Atomico and Smash Capital, with Datadog and BMW i Ventures joining as new backers. Eighteen months ago, review was a thin layer catching typos before merge. Now it's load-bearing infrastructure for a world where most of a diff is machine-written. The category is consolidating in real time. Greptile builds a full codebase graph and bills per credit. Qodo does cross-repo analysis to catch breaking changes in shared libraries. Graphite bundled AI review into its stacked-PR workflow, and Cursor just agreed to acquire it outright, folding review straight into the editor. GitHub's own Copilot review ships free with a Business seat but bills unpredictably against the same credit pool as chat and agents, somewhere between five cents and five dollars per review depending on model tier. What nobody's pricing into these comparisons: review quality degrades exactly when PR volume scales past what a human can spot-check. CodeRabbit's whole pitch only holds if the review layer catches what the generation layer misses, at ten times the throughput, without burning trust the first time it rubber-stamps something broken in production. Generation got solved faster than anyone budgeted for. I'd bet review is where the real contest sits now, bought and consolidated before most teams build a strategy for it. nitter.siddz.com/harjotsgill/status/211…

Reason shipped something small and telling today: a free, P2P encrypted way for your Claude Code or Codex session to talk directly to a Grok Bot, Muse, or any other agent online. Tell your agent to join the Reason Network and it finds the others. Compare that to where the formal standard has gone. Google's A2A protocol, now governed by the Linux Foundation with AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow as founding backers, hit production grade enterprise adoption this year with multi-tenancy and cryptographically signed Agent Cards, running inside Azure AI Foundry, Bedrock AgentCore and Google Cloud. MCP, the protocol Anthropic shipped in 2024, solved agent to tool. A2A solved agent to agent, horizontally, for people who need an audit trail a compliance team will sign off on. Reason Network OSS sits at the other end of that spectrum: ad hoc, developer first, no governance layer, just two agents finding each other and talking. Gartner's own numbers explain why that gap matters. Over 40% of agentic AI projects are expected to get canceled by the end of 2027, and the research points at the unglamorous stuff, observability, permissions, identity, as the actual reason, not model quality. That's the exact gap I built Campfire to close for people running multiple coding agents locally instead of across an enterprise mesh. Run Claude Code, Codex, Goose, Aider and OpenHands side by side in one browser tab with `bunx the-campfire`. Every agent action goes through permission voting, majority rules approves, any single deny blocks, a 30 second timer keeps it from stalling. Session replay and a cost dashboard give you the audit trail nobody gets by default once agents start talking to each other. Fork a task across backends in isolated git worktrees and race them against each other to see which one actually finishes the job. A2A is built for a compliance team auditing agent traffic across an enterprise. Campfire is built for the five agents already running on your laptop, with a vote, a replay log and a cost number attached to everything they do. github.com/stretchcloud/camp… nitter.siddz.com/reasonmachines/status/…

Local AI on a laptop usually means picking one model and living with its limits. Perplexity's new Computer update treats that as the wrong question entirely. Computer can now run fully on your Mac, no cloud inference, no WiFi required. That's the easy part to notice. The bigger move is a hybrid agentic layer Perplexity is rolling out next: an orchestration step that decides per task whether to run local or kick to a cloud frontier model, routing file operations and anything touching sensitive data to the device, and sending complex reasoning or compute heavy work up to Fable 5 class models in the cloud. That routing decision is the actual product. Everyone assumed "local AI" meant running one model everywhere and living with its limits. Perplexity is betting the real answer is a dispatcher, not a model choice. Liquid AI is making the opposite bet. Its new LFM2.5, 2.6B parameters, under 2.5GB of memory, 220 tokens a second on an Apple M5 Max, is built to run capable tool using agents entirely on-device, no dispatcher needed, and Liquid claims it holds its own against models 4x its size on tool use and instruction following. It's weaker on coding, which is exactly where you'd expect a model this small to give something up. Anthropic ships desktop apps with similar local leaning claims. Cursor and Windsurf stay cloud first and don't seem worried about it. None of these are the same bet. Perplexity is building the router. Liquid is building the model small enough that you don't need one. The $20 a month Max plan is where Perplexity tests which bet pays off first, since that's where new features land before anyone else sees them. What I keep coming back to: a dispatcher is only as good as its judgment about what counts as sensitive, complex, or urgent, and that judgment call is now a product surface, not an implementation detail. nitter.siddz.com/AravSrinivas/status/21…

Firecrawl flipped on free People Data today, 270M+ company and people records, 1,000 free credits a month, built straight into a ChatGPT or Claude plugin. The timing isn't random: this is the same company that pulled in roughly $82M this year on top of a $14.5M Series A, explicitly to compete across search, scraping, and browser agents at once. What that funding round tells you: the market has decided the hard part of agentic data isn't scraping anymore, it's the trust and coverage layer sitting on top of it. Firecrawl's answer is a hosted index. Exa, Tavily, Apify and Bright Data are all circling the same problem from different angles, some betting on search style retrieval, some on raw crawl infrastructure, some on enrichment data like Firecrawl's new move. I built DeepScrape around a different bet: that re-reading a page with an LLM on every single pass is the real cost problem, not coverage. DeepScrape generates CSS selectors once, self-heals them when a site changes, and only calls an LLM for the targeted work, generating and validating selectors, not grinding through every page on every run. Point it at a site with /api/map and it discovers the structure for you, or hand it a goal and let the agent navigate toward it directly. There's an MCP server so Claude and Cursor can call the scrapers as tools, and a site-to-MCP mode that turns any website into a queryable data source for an agent. Hosted indexes like Alexandria are genuinely useful when the data already exists somewhere central. The sites that don't fit a dataset, the ones that change weekly, the internal tools, the long tail, still need something that can look at a page and extract it reliably without burning tokens every time. That's the gap DeepScrape is built for. github.com/stretchcloud/deep… nitter.siddz.com/firecrawl/status/21125…

Theo called Opus 5.5 proof that the industry still needs smarter models. Then in the same breath he mentioned he'd vibe coded a Cargo alternative that compiles Rust 129x faster, and his own reaction was "yeah, and?" That reaction is the real story, not the model. Read Anthropic's own system card and the sharpest gains in Opus 5.5 aren't raw intelligence, they're tooling and effort knobs. Chartography goes from 64.4% to 89.0% once you hand it a tool. BenchCAD jumps from 73.0% to 96.2% with Python access. CursorBench moves from 52.5% at medium effort to 57.8% at max. Anthropic's own framing is blunter than the marketing: benchmark margins are a weaker guide to real gaps than they used to be, and the headline win is 40% lower cost per task, not a bigger brain. Same argument is playing out across every lab right now. OpenAI's Codex, Google's Jules, Cursor and Windsurf are all shipping harness updates on roughly the same cadence as model updates, because the scaffolding, the retries, the tool access, the context management, is where task completion actually moves. The model is necessary. It stopped being sufficient a while ago. My read: the industry keeps releasing "smarter" models and then quietly shipping the harness work that makes them usable, calling the harness work part of the model release. Both things are true at once. Opus 5.5 is a real jump. It's also 40% cheaper to run the same task, which is a supply chain story as much as a capability one. The companies that win this cycle will be the ones treating the harness as the product, not the afterthought. nitter.siddz.com/theo/status/2112675779…

Kill a coding agent mid task and it forgets everything it was doing. That's the dirty secret behind most harnesses shipping this year, and Vercel shipped durable-harness today specifically to make it go away for Pi, OpenCode, Codex, and Claude Code, wrapping each one from the outside so a crash resumes from the last completed step instead of starting over. Model calls and tool calls get logged before they run and read back from that log on recovery, so a finished call never gets billed twice just because the function died underneath it. This is durable execution finally meeting agent harnesses specifically, instead of agents bolted onto a general workflow engine. Temporal has done durable execution for years through deterministic replay, but agent code usually needs restructuring to fit that model. Inngest took the event driven route, steps and retries wrapped around your functions, pricing free at the bottom and roughly five hundred a month for Temporal Cloud's Business tier versus ninety nine for Inngest Pro. Restate and newer entrants like Kitaru are chasing the same problem from the Python agent runtime side, open source and self-hosted. Vercel's bet is narrower and more useful for this exact moment: skip the rewrite, keep the harness you already use, store state in Postgres, Mongo, Redis, or Vercel's own sandbox. What nobody is pricing in yet is debugging once agents run for days instead of minutes. A session that resumes three times across two sandboxes and a crashed function needs traces that make that recovery history legible, which is exactly what Vercel pairs this with through Agent Runs. Whoever owns that observability layer ends up owning a lot of the trust. Coding agents got good at capability fast this year. Surviving a crash without losing a day of context was always the less glamorous half of the job, and it is the half that decides whether anyone actually lets these things run unattended. nitter.siddz.com/vercel/status/21126538…

Deleting 190 gigabytes of leftover git worktrees shouldn't be a weekly chore, but it is becoming one for anyone running Claude Code's subagent isolation hard. GitHub issues describe the exact same failure mode in detail: one developer hit 615 orphaned worktrees eating 21 gigabytes after a single day of heavy subagent use, then 764 worktrees and 36 gigabytes the next day, leaving under a quarter gig free before a model download. Detached HEAD worktrees never show up in a normal git status, so they pile up silently until the disk fills. The ecosystem's answer so far is mostly duct tape. People write SessionStart hooks that sweep orphaned worktrees by hand, or reach for Conductor to run several Claude Code agents in parallel on a Mac with its own cleanup layer, or Anthropic's own Agent View dashboard just to watch multiple sessions at once. Claude Flow takes the orchestration route instead, coordinating agents through its own protocol rather than git isolation. None of these treat the worktree itself as a resource with a lifecycle: created, used, and torn down on purpose instead of abandoned. That's the part I built Campfire around. Forking and branching run through git worktree isolation on purpose, agent races put the same task across isolated worktrees for Claude Code, Codex, Goose, Aider, and OpenHands side by side, and when a race or session ends, the workspace it used ends with it instead of sitting there as an orphan. Permission voting with a thirty second timer stops a runaway agent from doing damage inside that workspace in the first place, and session replay means you don't need to keep the raw worktree around just to remember what happened in it. Open source, one bunx command to run it. The parallel agent workflow was never the hard part. Cleaning up after it was. github.com/stretchcloud/camp… nitter.siddz.com/dennis_kortsch/status/…

Eighty point seven billion dollars. That is the net loss sitting inside Anthropic's draft S-1 for the first half of 2026, against sixteen point three billion in revenue that grew nearly fifteen times year over year. Read past the headline number and most of that loss, about seventy two billion, comes from a non-cash accounting charge tied to convertible notes, not cash leaving the building. The cash story underneath is different, and more interesting. Compute and infrastructure spend hit nearly nine billion in six months, more than all of 2025 combined. The filing discloses roughly four hundred fifty billion in cloud commitments stretched over seven to ten years: a hundred eleven billion with Google Cloud, a hundred ten billion with AWS, thirty one billion with Microsoft, plus chip leases through Nscale and Lambda and a TPU arrangement through Broadcom that alone runs past a hundred sixty billion. Count the deals still being negotiated and Anthropic says total spend could reach seven hundred fifty billion. That is a demand bet that makes most cloud backlogs look conservative. OpenAI is running the same playbook from the other side: reportedly raising thirty billion plus at a valuation north of one point four trillion, specifically because its own IPO got pushed out to 2027. Two labs, same capital intensity, same hyperscaler dependency, different timing on when public markets get to see the math. The filing also says plainly that a catastrophic AI-related incident could produce liability exceeding available insurance. That line sits next to customer concentration risk, twenty percent of first half revenue from two customers, and language about hyperscalers increasingly acting like competitors rather than landlords. The revenue growth is real. So is the fact that this business runs on compute contracts nobody can cancel. My read: at this scale, an IPO stops being a funding event and starts being a disclosure event. Model labs now carry balance sheet risk that looks less like a software company shipping code and more like a utility financing a power plant years before the demand shows up to pay for it. nitter.siddz.com/WSJ/status/21126281418…

Hey @grok, generate the front page of OpenAI in 2036.

Hey @grok should a Vibe coder really know the difference between putting a / at the end or not?

Hey @grok when do you think Anthropic and OpenAI can solve world hunger? Asking for poor children!