Arun ArumugamAgent Engineering

Arun Arumugam · Agent Engineering

Managed by agents

OpenAI launched ChatGPT Wallet so agents can pay without seeing your card, llama.cpp reached a stable v1.0, and Google opened Antigravity's multi-agent /teamwork to outside researchers.

Latest

  1. NewsOpenAIOpenAI launches ChatGPT Wallet so agents can pay without seeing your cardBuilt with Amex, Mastercard and Visa: agents get tokenized cards and every purchase needs a security check. A concrete payments rail for agent checkout, starting with Pro users in the US.
  2. ReleaseDeepSeekDeepSeek API gets the app's search, files, voice, OCR, ASR and TTSBuilt-in tools switch on with a few parameters, and OCR, speech and TTS are also offered as standalone APIs, so developers get app-level features from one cheap provider.
  3. ReleaseHugging Facellama.cpp and ggml reach v1.0.0 with a stable API and ABIAfter 3.5 years, the core local-inference library commits to semantic versioning, so apps built on it can upgrade without surprise breakage.
  4. NewsGrok BotGrok Bot can now join meetingsThe bot sits in calls, answers questions, takes on assigned work and tracks team decisions, following Claude Meetings a day earlier into the same space.
  5. NewsGoogle ResearchGoogle opens Antigravity's /teamwork multi-agent tools after results on open problemsGoogle Research reports Gemini 4 multi-agent workflows producing results in math, formal verification, chemistry and kernel optimization, and is making /teamwork available to outside researchers.
  6. BlogCloudflareCloudflare sees 50+ AI pentest framework families scanning its customersForks of the withdrawn ARTEX agent now run from thousands of IPs. Cloudflare measures the traffic and explains how defenses must change for autonomous, high-volume attacks.
  7. BlogThinking MachinesHow Thinking Machines trained Inkling 2 Small end to end in 4-bitA detailed account of using NVFP4 in the forward pass from the first pretraining step through RL, and how they kept quantization error from hurting learning.
  8. BlogAnthropic ResearchAnthropic's 10 recommendations for building safety evaluationsPractical lessons on designing, running and reporting safety evals, since claims that a model is safer are only as sound as the evals behind them.
  9. BlogMeta EngineeringHow Meta brought Android's Cuttlefish emulator to Apple silicon MacsReality Labs ported Cuttlefish and its QEMU backend to macOS so AI-glasses developers can test against high-fidelity virtual devices instead of scarce hardware.
  10. ThemerisingSafety, security and provenance go operationalSafety and security work is turning into standing practice: OpenAI built Control Room so a monitor can halt any agentic research job in about 35 seconds, confidential compute keeps spreading from OpenAI's Private Inference API to Google's TPU 8tx, and Anthropic open-sourced FlashProve to prove a reply came from committed weights. The threat side is operational too, with Cloudflare now seeing 50+ families of AI pentest frameworks scanning its customers from thousands of IPs, while Anthropic published practical rules for building safety evals that hold up. Kill switches, sealed inference, provenance and sound evals are becoming day-to-day lab and network infrastructure.
  11. ThemerisingPersonal agents get their own computers, accounts and moneyAlways-on personal agents keep picking up real-world reach: Meta's Muse can send and receive email from its own address, Claude reads bank and card data through Plaid, and Grok Bot now works with Apple's Messages, Mail, Calendar and Contacts on the Mac. Money is the newest piece, as OpenAI's ChatGPT Wallet, built with Amex, Mastercard and Visa, gives agents tokenized cards with a security check on every purchase. Open options are arriving too, with n8n Labs' Nod and OpenCode Browse running agents on the user's own computer and browser. The agent is becoming a standing intermediary with its own identity, access and spending, so permissions and site-side controls matter as much as the model.
  12. ThemerisingOpen-weight models reach the frontierFrontier-scale open weights keep arriving: Thinking Machines put Inkling 2 Small, a 594B/17B MoE tuned for agents and coding, out under Apache 2.0 and then detailed how it trained the model end to end in 4-bit, after Mistral Large 4, Muse Spark Open, Reflection's Beam and Kimi K3.1 Pro. Open release is spreading beyond chat models with Runway's Praxis-1 for robot control, and the local stack is maturing as llama.cpp and ggml reached 1.0 with a stable API and ABI. Open release is a normal path to frontier capability, which tightens the gap for anyone building outside the closed APIs.
  13. ThemerisingAI turns to maths and science researchLabs keep pointing frontier models at research work: OpenAI published maths results and opened GPT-Rosalind to academic labs, Anthropic used Claude to help map a bacterial vault particle, DeepMind's AlphaGenome 2 traces variants to the cells where disease unfolds, and Google opened Antigravity's /teamwork multi-agent tools to outside researchers after Gemini 4 workflows produced results in maths, chemistry and kernel optimization. Formal methods are moving into real software, as AlphaProof Code verified security properties of Signal's protocol code, and OpenAI's BioBasicsBench scores agents on raw wet-lab data. Research results are becoming both a capability signal and a product line.
  14. ThemerisingAssistants move into the tools people already useAssistants keep absorbing the surfaces people already work in: Claude now builds whiteboards, live spreadsheets and data stories as native workspaces, and meetings became a contested surface within a day, as Claude Meetings and then Grok Bot started joining team calls to answer questions, take on work and track decisions. Muse Spaces turns a description into a persistent mini-app, and the flow also runs the other way, as OpenAI lets businesses deploy agents inside ChatGPT and Stripe exposes payments data to agents over MCP. Distribution through existing tools matters as much as raw model quality.
  15. SpikeNewsNVIDIAJensen Huang resigns as Nvidia CEO; CFO Colette Kress is interim CEONvidia's board says Huang has left as CEO and board member effective immediately, with Kress as interim CEO and Harvey Jones as the company's first board chair. Nvidia says its product roadmap and customer commitments continue unchanged, and Q3 results still land on Nov 17.
Show more
  1. ReleaseGoogle DeepMindGoogle DeepMind releases Gemini 4 Hydrogen, its fast everyday modelGoogle says Hydrogen finishes tasks 2.5x faster than Gemini 3.8 Flash and beats Argon on the Artificial Analysis Intelligence Index, with better instruction following and tool calls. Cursor reports 54.3% on CursorBench at default effort, up from 39.6% for 3.8 Flash.
  2. NewsAnthropicAnthropic launches Claude Meetings, Claude joins your team's video callsClaude can now sit in a team video call and build docs, whiteboards and other artifacts while people are still talking. It is rolling out in beta on Team and Enterprise plans.
  3. ReleaseThinking MachinesThinking Machines opens Inkling 2 Small weights under Apache 2.0The 594B-total, 17B-active MoE for agents and coding, launched last week, can now be downloaded from Hugging Face under a permissive Apache 2.0 license, so teams can self-host and fine-tune it.
  4. ReleaseWorld LabsWorld Labs opens its Atlas world model to everyone, with an APIAnyone can now use Atlas through World Labs' apps and API endpoints to turn an idea into a 3D world, then export keyframes, video clips or 3D Gaussian splats.
  5. NewsGrok BotGrok Bot on Mac can now use Messages, Mail, Notes, Calendar and ContactsThe personal agent's Mac app now works directly with Apple's built-in apps, including Messages, Mail, Notes, Reminders, Calendar, Location and Contacts, so it can act on your own data rather than only in the cloud.
  6. NewsAnthropicAnthropic launches its IPO roadshow with $17B quarterly revenueAnthropic opened its public offering with a prospectus showing $17.0B revenue in Q3 2026, more than 12x a year earlier, and over 11,900 customers paying $100K+ a year.
  7. ThemerisingClosed labs ship new flagships and faster tiersClosed labs keep shipping capability, speed and price tiers: Google opened Gemini 4 Carbon to all paying developers and followed with Gemini 4 Hydrogen, a fast everyday model it says is 2.5x faster than Gemini 3.8 Flash, while Anthropic's Claude Fable 5.2 is up to 30% faster and 20% cheaper at the top tier and Alibaba opened Qwen4-Max in preview. Cheap decision models are becoming their own category, with Google's Gemini Decision 1 scoring up to 1,000 questions per input in about 200 ms, while Vercel's AI Gateway shows small decision models already handle 26% of requests. Competition is shifting toward price per task and the workloads agents actually run, from flagship reasoning down to cheap high-volume calls.
  8. ThemesteadyAnthropic reshapes its governance on the way to an IPOAnthropic has opened its IPO roadshow, with a prospectus showing $17.0B of revenue in Q3 2026, more than 12x a year earlier, and over 11,900 customers paying $100K+ a year. It got there after filing an S-1 to list on Nasdaq as ANT, committing 5% of its shares to a Public Benefit Fund, and handing its Agent Skills standard to the Linux Foundation's Agentic AI Foundation. A frontier lab going public puts its finances and governance in the open and sets a template for how safety-focused labs raise capital.
  9. NewsQwenAlibaba opens Qwen4-Max in previewThe first Qwen4 model is usable now in preview on Alibaba's Token platform, pitched at coding, front-end builds and agentic work. It is a preview, so treat it as an early look before the full Qwen4 family lands.
  10. NewsCursorCursor rebuilds Projects so teammates' agents work togetherProjects is now multiplayer: every teammate and their agents share one workspace and can coordinate. It moves Cursor from one developer steering agents toward whole teams of agents working side by side.
  11. Newsn8nn8n Labs releases Nod, an open-source personal agentNod runs on your own computer with its own browser and works across apps and sites for chores like call prep, collecting invoices or cancelling subscriptions, with the AI provider you pick. It is an open alternative to hosted personal agents.
  12. ReleaseGoogleGoogle launches the Gemini Decisions API with Gemini Decision 1A model that scores up to 1,000 yes/no, choice or rating questions per input in one pass, in about 200 ms at $0.10 per 1M tokens. Google now joins OpenAI in selling decision models for agent workflows.
  13. Newsx.comGrok Bot 1.0 brings shared bots that work with whole teamsBots can now be shared across a team, work inside Slack and Teams, and join meetings. It moves personal agents into team workflows; the launch post passed 16M views.
  14. ReleaseRunwayRunway releases Praxis-1, an open-weight world action model for robotsApache 2.0 weights for a model that turns video pretraining into robot control, ranked first on NVIDIA's RoboLab-120. A strong open base for teams building robot agents.
  15. Newsx.comDeepMind's AlphaGenome 2 predicts how DNA variants act in specific cell typesIt models how each cell type reads the same genome, so researchers can trace a genetic variant to the cells where a disease actually unfolds.
Show more
  1. PaperPerplexityPerplexity red-teams its agent sandbox with 12 modelsModels tried to escape the VM and bypass network policy. Findings: standard capture-the-flag setups can steer agents away from the boundary being tested, and security-tuned models did not do better.
  2. BlogCloudflareWhy Cloudflare chose QuePaxa, an unusual consensus algorithmCloudflare's new Meerkat control plane uses QuePaxa, a leaderless consensus algorithm, because Raft and Paxos stall when a leader slows down on the open Internet. Explains the tradeoff and how it works.
  3. BlogMeta EngineeringMeta's Intelligent Delivery: making every unit of ads compute countMeta's system predicts the marginal value of compute per ad request and shifts serving resources to match, adding about 10% ranking performance. Now uses agents to detect and respond to system shifts.
  4. ThemerisingAgents become long-running, parallel and statefulAgents are moving from single chats to long jobs: parallel agents in ChatGPT Pro, Gemini agents playing Minecraft for 27 hours, durable pi agents and dedicated memory models, and GitHub rebuilding Copilot around parallel agent sessions. The harness is going multiplayer as well: Cursor rebuilt Projects so every teammate and their agents share one workspace, Cursor Cloud Agents scaffold and host live web apps, and Mistral Studio runs long-lived managed agents with approvals and tracing. Anthropic's write-up of agent memory treats offline "dreaming" as the pass that corrects bad notes and preloads the next session. The hard problems remain state, memory, orchestration and coordination around the model.
  5. ThemerisingSandboxes and limits for agents that actAs agents run code and touch production, containment is getting real engineering. Vercel's sandbox challenge surfaced a kernel zero-day, Codex on Windows moved to a container sandbox, Google offers a microVM per agent, and Figma built Nimbus to run tens of thousands of isolated sandboxes for untrusted code. Testing the walls is maturing too: Perplexity had 12 models try to escape its agent VM and found capture-the-flag setups can steer agents away from the boundary under test, and security-tuned models did no better. Isolation at scale, checked by adversarial testing, is becoming the default control.
  6. NewsGoogle DeepMindDeepMind's AlphaProof Code formally verifies Signal's protocol implementationGemini-powered agents proved key security properties of Signal's compiled protocol code in four weeks, writing 6M+ lines of Lean with humans reviewing about 1%. A first for a widely deployed protocol's implementation; artifacts are open-sourced.
  7. ReleaseGoogle DeepMindGemini 4 Carbon opens to all paying developers at Argon's priceTwo weeks after a cyber-defenders-only preview, Carbon is in the Gemini API at $2/$10 per M tokens until Nov 30 ($4/$20 after). Google says it tops CursorBench 4.0 and Zapier AutomationBench, and Antigravity previews /swarm multi-agent runs.
  8. NewsClaudeClaude adds Whiteboards, Sheets and DataClaude now builds boards, live spreadsheets and data stories as native workspaces, not just chat replies. It pushes Claude further into everyday office tools; rollout started today and the post drew over 4M views.
  9. ReleaseGoogle DeepMindGemini Robotics 2.5 hits 96% on a Mercedes-Benz assembly line pilotDeepMind's new vision-language-action model raises success rates 30-100% across 200 tasks and ran parts kitting on Apptronik humanoids in real factories. It also open-sources ASIMOV-Jailbreak, a robot-specific jailbreak benchmark.
  10. NewsOpenAIOpenAI launches Businesses in ChatGPT, with Business Agents in betaBusinesses can claim their profile in ChatGPT, see how people engage with them, and build an agent that answers customers inside ChatGPT. It turns ChatGPT into a discovery channel businesses must manage, like search listings.
  11. NewsMuseMeta's Muse gets its own email addressPaid Muse users in the US and Canada can now have their agent send and receive email on their behalf from its own address. Another big consumer agent getting its own identity rather than borrowing the user's inbox.
  12. BlogEngineering at MetaHow Meta built TLX, a warp-level extension to Triton for Blackwell and HopperMeta explains why neither CUDA nor Triton fit modern GPUs and how explicit warp-group orchestration got 20-25% MFU on ads training, up to 2/3 lower RecSys inference latency, and runs MoE pretraining kernels in production.
  13. SpikeReleaseAnthropicAnthropic releases Claude Fable 5.2 and Claude Mythos 5.2Anthropic's top tier gets faster (up to 30%) and 20% cheaper at $8/$40 per million tokens, with a 1M context. Fable is generally available; Mythos is the same model for verified cyber and life-science orgs.
  14. NewsGoogleGoogle adds TPU 8tx for confidential inference and TPU 8i Max for long contextGoogle now has four eighth-gen TPU systems: 8tx keeps data and weights encrypted in use for regulated agent workloads, and 8i Max doubles 8i's compute and HBM bandwidth. It also open-sourced TPU Sync and an agent suite for Claude Code and Cursor.
  15. NewsDeepSeek HarnessDeepSeek Harness reports 1.54M daily users and plans a plugin marketplaceUnder three months after launch, DeepSeek's open-source agent app says it serves 1.54M daily API users and 92T tokens a day, with 35,000 community plugins. The figures are self-reported; a plugin store and author support program are next.
Show more
  1. ReleaseMetaMeta releases Muse Ember, a small model for high-volume agent workBuilt for the tool calls, routing and coding subtasks agents repeat all day, at $0.10/$0.50 per million tokens with a 1M context. Meta claims it beats Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar.
  2. BlogOpenAI NewsHow OpenAI built Control Room to stop any research model workload in 35 secondsEvery agentic research job gets a job ID that inference engines check before generating, so a monitor can halt one job without killing shared engines. It fails closed: if Control Room is down, inference stops.
  3. BlogMeta EngineeringExtending LLDB to debug applications that span CPUs and GPUsMeta built a vendor-neutral accelerator plugin for lldb-server that shows host and GPU as separate processes in one session, so bugs crossing CPU code and kernels can be traced together.
  4. BlogNetflix Tech BlogFact-checking at scale: Netflix's synopsis factuality systemA real LLM-as-judge in production: Netflix works on cutting the cost of long-context reasoning calls and aligning the model's rationales with human annotators when checking synopses across titles and languages.
  5. BlogVercelVercel AI Gateway production index: decision models reach 26% of requestsMeasured from traffic of over 100 trillion tokens a month: small decision models grew to a quarter of requests while taking only half a cent of every dollar spent, and Anthropic is winning new teams 2:1.
  6. NewsOpenCodeOpenCode Browse lets any MCP agent use your real browserAn open-source MCP server and Chromium extension that lets OpenCode, Claude Code, Cursor or Codex drive your own browser with your logins, handing the tab back to you for 2FA or payments, with per-session site allowlists.
  7. BlogLangChainHow LangChain optimized browser use for Deep AgentsA measured comparison of seven browser toolsets: code-mode Stagehand passed 90% of WebVoyager for about $2 versus 78% and $27 for Playwright MCP, plus a trace-driven fix to agent sleeps that cut median task time by a third.
  8. NewsMetaMeta launches Spaces: custom mini-apps Muse builds from a descriptionDescribe a planner, tracker or hub and Muse builds it into the app, with tabs you can pin or schedule and connected apps that keep it updated in the background. Available everywhere Muse is.
  9. NewsOpenAIOpenAI introduces BioBasicsBench for AI agents on wet-lab analysis100 tasks from real Ligo experiments test whether agents can read raw gels, plates and instrument data, catch errors and refuse unsupported conclusions. GPT-6 Astra scores 66.7% at xhigh vs 43.8% for GPT-5.6 Sol.
  10. BlogAnthropic EngineeringAnthropic on agent memory: dreaming to correct, organize and preload notesEngineering write-up of how Claude keeps experience across sessions, and how offline "dreaming" fixes bad notes, builds wiki-style memory and decides what to load into the next session.
  11. SpikeNewsclaude.comClaude adds Personal Finance: read-only bank and card connections via PlaidUS Pro/Max/Ultra beta lets Claude answer money questions from balances and up to two years of transactions you connect; access is read-only and hub chats are not used for training.
  12. SpikeReleaseThinking MachinesThinking Machines ships Inkling 2 Small, a 594B/17B MoE for agents and codingSecond-generation Inkling scores 31 on the AA Intelligence Index; available now on Tinker for inference and fine-tuning, with Apache 2.0 weights planned next week after safety checks.
  13. Newsopenai.comOpenAI's autonomous pipeline found and patched 948 Linux kernel privilege-escalation bugs in one runAn upcoming OpenAI cyber model, run in an autonomous pipeline, found, validated and patched 948 Linux kernel privilege-escalation bugs in a single run, a concrete marker for security agents.
  14. Releaseopenai.comOpenAI ships ChatGPT Images 2.6 and two faster GPT-Image-2.6 API modelsChatGPT Images 2.6 adds GIFs and reference images, and the new API models Flare and Sunburst generate images about 50% faster, with Sunburst keeping details more consistent across edits. They're also available in Codex.
  15. Newsperplexity.aiPerplexity launches Teammates, persistent agents that own ongoing work inside ComputerMax-tier teammates plan and follow through on long-running jobs using Computer tasks, skills, memory and tools, can collaborate with other teammates and message coworkers in Slack.
Show more
  1. Newsabout.fb.comMeta rolls out Muse Voice and Realtime Avatar in the USUS users can talk to Muse in real time on mobile, web and Mac, with expressive avatars; voice is coming next to AI Glasses and Muse Charm.
  2. Newscursor.comCursor Cloud Agents can build, host and share live web apps from a promptDescribe a tool and Cursor scaffolds it, provisions Postgres if needed, runs security checks, and returns a shareable hosted URL behind Cursor auth or a public link.
  3. Newsmistral.aiMistral Studio adds managed agents and apps for durable enterprise agent deploymentsStudio now runs long-lived managed agents with approvals, sandboxes and tracing, plus one-command app deploy, on shared infrastructure with Le Chonk.
  4. Newsblog.cloudflare.comCloudflare adds personal-agent analytics and granular access policies for sitesSite owners can see which verified personal agents hit which paths on their behalf and set finer allow/block rules as agents become a first-class traffic class.
  5. Newsstripe.comStripe lets any business analyze payments data through an agent via MCPNatural-language queries over Stripe data for reports, fraud and retention metrics are generally available; agents connect through the Stripe MCP or Queries API.
  6. BlogGitHub EngineeringInside GitHub's new Git storage architectureA detailed write-up of how GitHub redesigned Git object storage for scale, with the failure modes and tradeoffs that drove the new design.
  7. BlogFigmaHow Figma built Nimbus to run tens of thousands of AI sandboxesFigma's AI products need many isolated sandboxes for untrusted code; Nimbus is the internal platform that makes them fast, reliable and operable at that scale.
  8. BlogCloudflareCloudflare on why sites must treat personal agents as a new intermediary classA clear framing of how personal agents strip user identity from requests and a collaborative approach for sites to identify, allow and manage that traffic safely.
  9. ThemefadingInfrastructure strains under agent-scale codeAgents now write code faster than Git hosting and review were built for. GitHub moved repositories to Azure under agent-driven load and has published the redesign of its Git object storage; OpenAI built Pando, a Rust Git backend for thousands of concurrent pushes from engineers and Codex agents, cutting median push latency by about 80% on a petabyte-scale monorepo. Anthropic built its own source control after merges grew more than tenfold, and teams are answering review load with automation such as Meta's just-in-time tests. Tooling built for human pace is giving way first.
  10. ReleaseTencent HunyuanTencent releases Hy4, its new flagship model with vision and computer useTencent's Hunyuan team shipped Hy4, adding vision and computer use for the first time, and says it ranks among the leading open models. It's available on Tencent Cloud and OpenRouter, so builders can try it today.
  11. NewsAnthropicAnthropic used Claude to help map the first 3D structure of a bacterial vault particleA concrete case of a frontier model contributing to real structural biology work, from Anthropic's own research team.
  12. NewsGrok BotGrok Bot can now make phone callsPersonal agents are moving beyond chat and apps into voice calls placed on the user's behalf.
  13. PaperAnthropicAnthropic open-sources FlashProve, zero-knowledge proofs that a reply came from a committed modelProviders commit to hidden weights and prove each reply came from that model; a ~19 MB proof checks on a CPU in seconds. Runs a 400B MoE at ~110 tok/s on one H100, soundness proved in Lean. A step toward verifying the model you pay for.
  14. BlogOpenAIOpenAI built Pando, a new Git backend for its agent-heavy monorepoRust, Git-compatible backend for thousands of concurrent pushes from engineers and Codex agents: per-branch coordination instead of repo-wide consistency, and packfiles reused across fetches from regional caches. Median push latency down ~80%; ~1 PB of Git data served daily.
  15. NewsOpenAIOpenAI introduces LifeVal, a benchmark for everyday personal-assistant tasksNew eval of real-life tasks (childcare, groceries, government paperwork) done via browser, computer use and connectors. GPT-6 Astra scores 80.6% vs 49.3% for GPT-5.2. Self-reported; no task count, grader details or other labs.

Older onPostsThemesStartupsFeed

Startups

  1. StartupAll startupsLangChainLangSmith, a model- and cloud-neutral platform to build, run, test, monitor and govern AI agents, on top of the open-source LangChain, LangGraph and Deep Agents frameworks; 7,000+ organizations use LangSmith.
  2. StartupAll startupsAgile RobotsMunich maker of humanoid and mobile robots, robot arms and hands, and the AI models behind them, for factories and warehouses; more than 20,000 robotic solutions installed.
  3. StartupAll startupsMirendilTools that let one researcher run hundreds of AI agents writing code and running experiments for days, to put frontier AI R&D within reach of small teams; trained its own coding model, Fern.