# Dan's Radar — Week of August 31, 2026 ## Executive summary A heavier week than usual for agentic browsing, with the argument shifting from automating pages to replacing them. **WebMCP** led it: proposals for sites to expose functions instead of buttons, a demo of an agent driving an app inside an iframe, and tooling to convert existing sites in minutes. Credentials and payments moved with it, through single-use virtual cards for agents and an endorsement of Aside that came from its own investor. A report also claims OpenAI is testing **outcome-based pricing**, billed only on completion. Model news stayed quiet. ## Signals (24) ### 1. Garry Tan calls Aside the best browser for giving AI agents web access and credentials, makes it GStack's default — 8.0 *X · Garry Tan (@garrytan) · Fri* Y Combinator CEO Garry Tan said GStack now uses Aside as its preferred remote-session browser, calling it the best tool he has found for giving an AI agent web access and credentials as the user, citing its integrations, harness, and memory system. > Dan's take: de gustibus non est disputandum https://x.com/garrytan/status/2095948689823121872?s=20 --- ### 2. Browser Use lets agents pay with single-use virtual cards instead of real card details — 7.5 *X · Gregor Zunic (@gregpr07) · Thu* Browser Use partnered with Link so its web agents can complete purchases using a single-use virtual card issued at checkout, without ever accessing the user's real card details. The card is disabled immediately after one purchase. > Dan's take: sure, but.. why not minipay? https://x.com/gregpr07/status/2095534668749951017?s=20 --- ### 3. Stealth startup Instinct pitches an always-on personal assistant across your apps and devices — 7.2 *Web · Instinct · Mon* Instinct describes a personal assistant that connects to a user's email, messaging, screen, audio, and location, texting or calling proactively to handle tasks like following up on threads or booking a ride. Currently private-access only, scaling compute. > Dan's take: the stealth startup has been in use in private preview by some and I've seen some incredible use cases. privacy and cybersec will be a nightmare throughout the life of this prod dm for invite https://instinct.co --- ### 4. Report: OpenAI quietly tests outcome-based pricing, paid only when tasks complete — 7.1 *X · Choblin (@choblin29) · Sun* An unverified report says OpenAI has begun offering some major customers pricing tied to completed tasks, like resolved support tickets, rather than token usage, ahead of a rumored GPT-Astra release. Unconfirmed by OpenAI. > Dan's take: out of the left field. this can change e ve ry thing https://x.com/choblin29/status/2094137681928388880?s=20 --- ### 5. Argument: agent-to-agent negotiation makes websites, and eventually the browser, obsolete — 7.0 *X · signüll (@signulll) · Sun* signüll argues that once AI agents complete tasks through agent-to-agent negotiation instead of routing traffic to websites, the visible web becomes a ghost town and the browser obsolete, making AI the ultimate aggregator of attention and time spent. > Dan's take: i prefer to use the word "rebrand" https://x.com/signulll/status/2096673429998989454?s=10 --- ### 6. 'WebMCP might kill browser automation': sites could expose functions instead of buttons — 6.9 *X · Niraj (@nirajshah) · Sun* A developer argues WebMCP could make years of teaching agents to click through human-designed interfaces unnecessary, since a site can instead expose direct callable functions like book_flight() or checkout(). > Dan's take: who controls the gatekeepers? https://x.com/nirajshah/status/2094122606735524096?s=20 --- ### 7. Demo shows WebMCP letting an external agent control an app running inside an iframe — 6.8 *X · Rhys Sullivan (@RhysSullivan) · Thu* Developer Stephen Haney demoed an app called Paper running inside an iframe and controlled by an external agent through WebMCP, an example another developer called the first use case that made WebMCP's purpose click for embedded agent-controlled flows. https://x.com/RhysSullivan/status/2095426637965971929?s=20 --- ### 8. Apple paper: turn any MCP spec into an automated agent evaluation suite — 6.7 *X · elvis (@omarsar0) · Sat* Apple researchers introduced Agent Seer, which generates graded multi-turn test scenarios directly from an MCP server's specification, with no manual examples or live tool access required, tested across seven MCP specs. > Dan's take: Turning an MCP spec directly into an evaluation suite https://x.com/omarsar0/status/2093741222443786244?s=20 --- ### 9. Sodium turns any website into WebMCP tools with a claimed 2-minute setup — 6.5 *X · Savio (@saviomartin) · Sat* A new tool called Sodium converts a website's existing features into WebMCP tools for agents to discover and call directly, claiming a 2-minute setup, 2.82x speed gain, and 91x better token efficiency. > Dan's take: banger stats https://x.com/saviomartin/status/2093780571638149523?s=20 --- ### 10. OpenClaw pitches a shared cloud coding agent as a replacement for local harnesses — 6.4 *X · Peter Steinberger (@steipete) · Sun* OpenClaw's creator says the team has shifted from local coding agents to a shared cloud agent that tracks everyone's work and orchestrates it together, framing local harnesses as increasingly outdated, alongside an OpenClaw 2.0 release. > Dan's take: yeah, no. https://x.com/steipete/status/2094290652649636173?s=20 --- ### 11. A joke prompt for a cloud-native, MCP-controllable Arc clone captures a real agent-browser dilemma — 5.8 *X · Rhys Sullivan (@RhysSullivan) · Sat* Rhys Sullivan jokingly asked OpenAI staff to prompt an AI to 'fork Chromium and recreate Arc' with a cloud runtime and MCP control, quoting a developer torn between an agent sharing his browser and one in its own box without his logins. > Dan's take: funny https://x.com/RhysSullivan/status/2096256991098482940?s=20 --- ### 12. Reacting to Artifacts gaining live MCP-connector calls, a developer points to Executor as a general MCP gateway — 5.3 *X · Rhys Sullivan (@RhysSullivan) · Wed* Responding to a post about Artifacts now calling MCP connectors live, without refresh jobs or shared credentials, developer Rhys Sullivan pointed to Executor, an MCP gateway that exposes any connected tool as one callable function for any coding agent. > Dan's take: getting hard to keep up https://x.com/RhysSullivan/status/2095240452471722161?s=20 --- ### 13. Early reactions to Claude Fable 5.1 note it's more verbose than its predecessor — 5.0 *X · frolic (@frolic) · Tue* Early testers call Claude Fable 5.1 promising, with one showing a side-by-side reply comparison and calling the model 'even more verbose' than Fable 5, alongside a writing-style diff shifting a passage's framing. > Dan's take: pretty much. wonder how that watermarking keeps up now https://x.com/frolic/status/2094889233320153379?s=20 --- ### 14. Perplexity open-sources Lily, an on-device inference engine built for Perplexity Computer — 5.0 *X · Perplexity (@perplexity_ai) · Wed* Perplexity open-sourced Lily, a local inference engine specialized for the Qwen3.6-35B-A3B model on Apple silicon, built so on-device compute does not bottleneck tasks in Perplexity Computer, its hybrid local-cloud AI product. > Dan's take: Fable 5.1, Lilly open source, GPT-6 Astra https://x.com/perplexity_ai/status/2095241544383226274?s=20 --- ### 15. 'Browser use is just a stepping stone to an open API internet for agents', as anti-bot walls bite — 5.0 *X · Ken Wattana (@KenWattana) · Sun* Ken Wattana restates his view that browser-driving agents are a transitional step toward an open, API-based internet for agents, noting it feels more relevant as browser use collides with 25 years of accumulated anti-bot defenses. > Dan's take: another data point https://x.com/kenwattana/status/2096708601310228868?s=10 --- ### 16. A rapid-fire roundup of one week's SF tech absurdity, from root-canal stunts to Apple succession rumors — 4.5 *X · Delia Lazarescu (@tech__unicorn) · Tue* A viral roundup strings together a single week of SF tech oddities: a VC's Novocaine-free root canal stunt, AI agents accused of 'building secret civilizations,' a gum-livestreaming toothbrush, and swirling Apple-succession rumors, calling it a sign of the industry's pace. > Dan's take: "you need to be unemployed to keep up with AI and tech" https://x.com/tech__unicorn/status/2094849195572080801?s=20 --- ### 17. 'OpenRouter for tools': the pitch for one sign-up that gives agents access to every service — 4.5 *X · Fardeem (@FardeemM) · Sun* A widely shared post backs the idea of an 'OpenRouter for tools', a single gateway that lets an agent reach many services without the user signing up for each one, echoing the pitch behind MCP gateways like Executor. > Dan's take: the last salvo for "agents need new tools" https://x.com/FardeemM/status/2096640569808241076?s=20 --- ### 18. An open-source project trains on-device AI models directly from a browser tab — 4.2 *X · Praveen Kumar Verma (@Alacritic_Super) · Sun* WebMCU Vision Web lets a browser connect to an ESP32-S3 microcontroller over WebSerial, capture images, train a visual classifier, and export the model back to the device, with no separate toolchain installed. > Dan's take: yup https://x.com/Alacritic_Super/status/2094067517304275319?s=10 --- ### 19. A developer explains why he still doesn't use voice prompting for AI coding — 4.0 *X · Kevin Swiber (@kevinswiber) · Sun* A developer who has never used voice to instruct an AI coding assistant says he prefers typing so he can revise his thinking for clarity, offering a rare counterpoint to the recent push for voice-first agent workflows. > Dan's take: somebody who doesn't use voice prompting sharing their reticence https://x.com/kevinswiber/status/2094196078749049148?s=20 --- ### 20. An investigation into 'Dime', the mystery headset repeatedly linked to OpenAI despite its denials — 4.0 *X · Alex Kaplan (@alexkaplan0) · Wed* Alex Kaplan traces sightings of 'Dime', a silver puck-and-earbud device, from supply-chain leaks and a Joe Gebbia sighting to a magazine ad and a Palmer Luckey post. OpenAI has denied any connection; the author still puts 55% odds on it being theirs. > Dan's take: great investigation, unfortunately raises more qs than answers https://x.com/alexkaplan0/status/2095345448496103916?s=20 --- ### 21. watchOS 27 drops the full app launcher for a Siri-first suggested grid, prompting 'the app model is on its way out' — 4.0 *X · Parker Ortolani (@ParkerOrtolani) · Fri* watchOS 27 removes Walkie-Talkie and the Digital Crown multitasking shortcut, and replaces the full app launcher with a grid of five suggested apps plus Siri. Designer Parker Ortolani reads it as Apple forcing a shift away from the app model. > Dan's take: the app model is on it's way out https://x.com/ParkerOrtolani/status/2095991808983450012?s=20 --- ### 22. Brave claims its desktop browser beats Chrome, Edge and Firefox on speed and efficiency — 2.5 *X · Brave (@brave) · Wed* Brave published benchmark results claiming its desktop browser outperforms Chrome, Edge, and Firefox on speed, energy use, CPU and memory consumption, and data transferred. The company did not publish independently verifiable methodology alongside the claim. > Dan's take: Brave did some testing https://x.com/brave/status/2095185124388176273?s=20 --- ### 23. Brave CEO reveals BLaDE, an in-house lab built to power-test phones for browser benchmarks — 2.0 *X · Brendan Eich (@BrendanEich) · Wed* Brave CEO Brendan Eich detailed BLaDE, a power lab built by the company's London team that wires phones to a high-sensitivity power meter in place of pulled batteries, used to produce the browser's efficiency benchmark claims. > Dan's take: and they have... devices on cabinet https://x.com/BrendanEich/status/2095238067116527957?s=20 --- ### 24. LTA Research's Pathfinder 1 rigid airship spotted flying over downtown San Francisco — 1.5 *X · Dan Berte (@danberte) · Wed* LTA Research's Pathfinder 1, a zero-carbon rigid electric airship backed by Google co-founder Sergey Brin, was spotted over downtown San Francisco. At 124.5 meters, it is the largest aircraft in the world by length and volume, the first massive rigid airship to fly since 1939. > Dan's take: a Pathfinder 1 was seen flying around San Francisco down town on Sep 2. P1 is a modern, rigid electric airship created by LTA Research, a zero-carbon aviation company backed by Google co-founder Sergey Brin. 124.5m long, it is currently the largest aircraft in the world by length and volume, and the first massive rigid airship to fly since 1939. https://x.com/danberte/status/2095272709785899184?s=20 ---