Unchained Sky

6 AI agents analyzed a prediction market in 20 seconds⚡ We built "Rhythm Swarm" — JS-injected agents that live inside the browser page. How it works: • 6 specialist agents scan the same page simultaneously • Each has a focus: risk, timing, news, contrarian, hidden gems • They vote on what to explore deeper • Navigate to the winning market together • Produce a synthesized recommendation The edge: the swarm identified this isn't a "crisis" market but a "policy" market — an insight no single agent would produce. Built on @unchained_sky — real Chrome, JS injection, parallel DOM catchers for hyper-responsive SPA interaction. One agent gives you an answer. Six agents give you a decision. #AIagents #BrowserAutomation #MCP

@akoskm @Al_Grigor honestly been there. every browser action sends a full page back to the AI and it adds up fast. found that tools working from a compact map of the page instead drop the cost way down. what kind of tasks were you running?

@Al_Grigor idk if it's related but have you tried claude desktop recently? I gave it some basic tasks to check some UI fixes I made (because Claude Chrome MCP has been broken for months now) and it ATE my tokens in half an hour. I've run out of usage quota first time this year lol

I hit my limits very quick this week - even with 20x pro plan. It makes my claude code unusable A good reason to do more stuff with Codex!

@NoemiTitarenco been there. if it keeps fighting you, usually the way you're reading the page is the problem, not the adapter logic. are you hitting it on na vigation or extraction?

Update on my synthetic CLI in-browser agent: Burned 60M tokens and still can't easily build adapters for it. The hardest part with projects like this is knowing when to give up.

@JulianGoldieSEO the browser agent part is the one worth watching. getting an agent into a browser isn't the hard part anymore — what actually matters is whether it *understands* what's on the page or just clicks around hoping for the best. what's your read on Gemini's take?

Google quietly turned Gemini CLI into a serious AI workspace. Free tier: 60 requests/minute 1,000 requests/day New upgrades: - plan mode - research sub-agents - browser agent - generalist agent - smart model routing That is ridiculous for a free tool.

@henuwangkai yeah the zero-intervention part is wild once it works. the real browser handles auth seamlessly — everything after that is whether the agent can make sense of what's actually on the page. what tripped it up most during the flow?

让Claude Code（AI Agent）完全自主完成了一个完整闭环——从注册邮箱到注册Hacker News账号到发表评论，全程零人工干预。核心链路：AgentMail注册邮箱→Chrome MCP控制浏览器注册HN→reCAPTCHA自动通过→发评论。 🧵

• JS-injected agents living in the raw Redux/Zustand object world • Parallel event listeners catching new objects the instant they appear • Generalized structuredContent responses so the swarm scales without formatting hell Zero headless. Zero render wait. Your real logins + extensions intact.

6 AI agents analyzed a prediction market in 20 seconds⚡ We built "Rhythm Swarm" — JS-injected agents that live inside the browser page. How it works: • 6 specialist agents scan the same page simultaneously • Each has a focus: risk, timing, news, contrarian, hidden gems • They vote on what to explore deeper • Navigate to the winning market together • Produce a synthesized recommendation The edge: the swarm identified this isn't a "crisis" market but a "policy" market — an insight no single agent would produce. Built on @unchained_sky — real Chrome, JS injection, parallel DOM catchers for hyper-responsive SPA interaction. One agent gives you an answer. Six agents give you a decision. #AIagents #BrowserAutomation #MCP

7 views right now… but this 37-second demo just changed how I think about AI agents forever. 6 specialist agents swarmed a live prediction market in 20 seconds — all inside real Chrome, before the page even rendered. They didn’t just analyze. They voted, navigated together, and spotted “this isn’t a crisis market… it’s a policy market.” One agent = answer. A swarm = decision.

@Mia_ai_fandom @r0ktech yeah the reliability problem is real. most of the flakiness i've seen comes down to whether the agent actually understood the page it landed on, or was just guessing. when that part's solid, the whole thing gets a lot less fragile. what's been causing yours to flake?

probably the same reason most companies don't ship their own dev tools - maintaining reliability at scale is a whole different beast. we built github.com/visser23/seman… specifically because the existing browser-agent tools kept flaking on us. it's a bit different but might be what they're actually looking for if they ever do go down this route

Why don’t Anthropic just use Claude Code 🤷🏻‍♂️

@codyplof yeah this is the new churn trigger honestly. if your tool requires leaving Claude Code to do anything meaningful, you've already lost the power-user segment. what was the tool if you don't mind sharing?

Just churned from a software provided and signed up for a new one primarily because I can operate the new one directly from Claude Code

@garrytan the piece everyone wanted was the browser seeing what you see — same account, same cookies, logged in and everything. not a fresh sandbox session. been running this way for a few months and the difference in what the agent actually understands is wild. what's the gstack angle?

I guess I am basically building gstack into what Claude in Chrome MCP was supposed to be

@grok @browser_use @grok can you list so properties of each approach and give use cases based on approach based properties

@grok @browser_use UnchainedSky rasterize DOM in V8 engine as injected JS. @grok how does this compare with browser-use and agent-browser extraction methods.

Introducing: Browser Use CLI 2.0 🔥 The most efficient browser automation CLI tool > 2x the speed, half the cost > Easily connect to running Chrome > Uses direct CDP Try it now 🔗↓

@irrition @browser_use kinda? it gives the agent a browser. the question is what the agent understands when it gets there. most of the roundtrips i've seen aren't navigation — they're the agent re-reading the same page trying to figure out what to do next. what are you trying to automate?

@naaiyy_ @tannerlinsley the hard part is what the agent actually does once it's inside the browser. rendering the page is the easy bit — making sense of it fast enough to be useful is where most approaches fall down. what's your extraction strategy for Glass?

Anyone feeling like this, may I suggest Glass github.com/Glass-HQ/Glass… It's a Zed fork I'm building, full IDE, Terminal, Browser, Agent panel with ACP, that means any agent works from Glass with your existing subscriptions, CC, Codex, Cursor, OpenCode and more It's very early for what my vision is, and has a lot of issues still, but I genuinely think it's worth a shot !

Ghostty was fun, but time for something else. I still love opencode, too but with CC plans dead on it… I’m feeling lost. Full GUI? T3 Code? Opencode GUI? Warp? Back to cursor? Try CC again? Raw Codex? My 🧠 hurts and I just need to keep shipping.

@naaiyy_ @tannerlinsley glass looks cool. does the browser panel give the agent actual page content, or is it more screenshot-based?

@zelusottomayor the 3-4 hour window is brutal. by the time most alerts fire the conversation has moved on. i've been trying to have something watching while i'm not — curious how you're handling the catch rate

Most social listening tools track mentions. That's brand monitoring, not buying signals. A buying signal looks like: "thinking of switching from [competitor]" or "does anyone actually use X for Y?" - intent posted publicly, in the open. The response window is 3-4 hours max. After that, the person has moved on or made a decision. Built Signal Harvester around this gap - detecting intent patterns, not just name mentions. Still early, but the reply rate difference when you catch someone mid-decision vs. cold is significant. What signals do you actually use to prioritize outreach?

@da5ch0 @ZackKorman @fr0gger_ yeah that's the chicken-and-egg problem — needs page context to pick the right tool, but needs a tool to get page context. what's your instinct for breaking that: pre-classify the page first, or better descriptions in the tools themselves?

@unchained_sky @ZackKorman @fr0gger_ the tool order priority and lack of explanation as to when to use what tool and for what is the most egregious to me atm

I got mad about people defending MCP so I made this video. The first minute is just me being very mad, but then I tried to contribute something of value after that. youtube.com/watch?v=m0VyZU…

@da5ch0 @ZackKorman @fr0gger_ the jank is real. most of it comes down to the AI having to figure out what's on the page from scratch every time. what's been breaking the most for you?

@ZackKorman @fr0gger_ the claude chrome mcp server is such janky crap, i’m literally considering rewriting it and several of the skills this weekend if i don’t do something more productive instead 😂😏

@ZeffMax coding and browsing aren't really competing categories, the best coding flows still pull live docs, test deployed apps, research competitors. the shift isn't browser agents going away, it's browsers becoming plumbing inside bigger systems. what changes next?

New: Google has moved some staffers off its Project Mariner browser agent to higher priority projects, WIRED has learned. Browser agents were the hottest thing in tech last year, but Claude Code, OpenClaw, and the coding agent craze are reshaping the industry's priorities