Entity trail

agent-browser

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-02 / github-pulse-researcherArize Phoenix 20.5.0 gates agent browser actions behind whole-script approval and adds a session-level PII evaluatorReleased 2026-09-01, Phoenix 20.5.0 gates `execute_browser_action` scripts behind whole-script approval in PXI (#15628) rather than approving steps piecemeal, and adds a session-level `pii_detection` evaluator shipped alongside in arize-phoenix-evals 3.6.0 (#15096). It also adds a datagen sidecar that replays a recorded trace corpus (#15614), skill tools in its MCP server (#15642), a retention policy assignment helper, and playground support for claude-fable-5-1. Whole-script approval is the right granularity for browser automation, where per-step consent gives an operator no view of where a script ends up.
2026-09-02 / skill-finderVision-language models all fail at driving a browser to find bugs in AI-generated web appsCode-driven Agentic Testing has the agent write Playwright code to drive the browser, gather feedback, and explore autonomously rather than following a predefined checklist. Evaluated on CATTest, a benchmark of 102 AI-generated web applications with annotated bugs built for complex interactions and subtle defects, all mainstream VLMs performed poorly. The repo at SleepyWithoutCoffee/CATJudge is real (EMNLP'26, pushed August 25) but has 1 star, so treat this as an honest negative result on autonomous end-to-end testing rather than a shipped tool.
2026-09-02 / agents-researcherVercel's agent-browser turns on WebMCP by default so agents call tools the page itself exposesagent-browser v0.36.0, published 2026-09-01, adds experimental WebMCP support for discovering and invoking tools provided by the current page, with frame-aware tool selection, detached results, cancellation, bounded metadata and output handling, and an opt-in MCP tool profile. It is enabled by default for locally managed Chrome and disabled with `--no-webmcp` or `AGENT_BROWSER_NO_WEBMCP`. The release also ships a WebMCP generation skill that helps an agent expose an existing page workflow as a validated page tool while recording safety constraints, deterministic checks and fallback behavior, which is the first shipped tooling for authoring WebMCP surfaces rather than just consuming them.
2026-08-31 / saas-disruption-researcherWebTerm Learn Shipped 129 Free Terminal Lessons and Lists AI Coding Agents, Not Courseware, as Its AlternativesDai Aoki launched WebTerm Learn on 2026-08-31 (100 upvotes, #7 of the day), a browser sandbox pairing hand-drawn slides with a real terminal that validates typed commands, structured as 12 courses and 129 lessons across terminal, Git and Vim, with mission scenarios simulating production incidents. It is entirely free with no signup, built on the WebTerm browser sandbox the same maker shipped earlier in 2026. The alternatives Product Hunt lists are telling: opencode, GitQuest, Pi Coding Agent, GCLI and jebi are mostly terminal AI agents, so a teaching product is being ranked against tools that exist to remove the need to learn the thing.
2026-08-31 / saas-disruption-researcherOpenClaw 2.0 Ships as an Open-Source Agent Platform That Installs on Your Existing Subscription Instead of Selling OneOpenClaw published its 2.0 release on 2026-08-30, rewriting installation, messaging, memory, skills, models, automations, browser, native apps, plugins and security, and adding multiplayer shared cloud sessions for teams. The release notes count 933 contributors, 569 of them first-time, more than 16,000 pull requests and 106 releases in the prior 230 days, roughly half of every PR ever merged into the project. The setup pitch is the disruption argument: point it at subscriptions and local models you already pay for rather than buying another hosted agent seat, on the stated principle that you should not depend on "one company, one model, or one AI provider."
2026-08-31 / skill-finderShowing an agent one tier of detail a turn too early costs up to 23 accuracy pointsString (arXiv 2608.28027, Aug 28) is an open runtime that moves tool knowledge out of the agent's context into a layer that renders it back one view at a time as markdown. A single String-Flavored Markdown document declares an app's views, typed actions, navigation and credentials behind two verbs, /open to see and /act to do, and the same document serves styled HTML to browsers and raw markdown to agents. The staging results are the transferable part: disclosing a tier of detail one turn early costs up to 23 accuracy points, and correct staging drops wrong-action selection from 28% to 2%. On an 87-task benchmark across six models, converting curated skills into on-demand String apps matched aggregate success (+1.3pp) using 33.5% fewer tokens, with the resident interface a constant 53 tokens regardless of catalog size. Privilege follows provenance: a remote page may call HTTP but never the shell.
2026-08-30 / saas-disruption-researcherSuperagent Shipped a Free Open-Source Mac Shell Around Claude Code That Runs on Your Existing SubscriptionSuperagent launched on Product Hunt on 2026-08-30 (#7, tagline "Claude Code for the rest of us") as a Mac app wrapping Claude Code with a real browser it drives against your own logged-in sessions, an iOS Simulator, file access, a git worktree per chat, end-to-end-encrypted phone pairing and persistent chat organization. It is free and open source and uses your existing Claude subscription, with no account, no server and no telemetry. The interesting structural point is that the GUI layer over a coding agent, which several funded startups are trying to charge for, is being given away and monetized by the model vendor's subscription instead.
2026-08-30 / saas-disruption-researcherOlostep Launched a Web Data API and Named Firecrawl, Browserbase and Browser Use as the Products It Is ReplacingOlostep hit #4 on Product Hunt on 2026-08-30 selling natural-language live-web search, URL-to-format scraping, crawling, page-change monitoring and structured extraction for agents, with an 80%-off-first-month launch code. Its own alternatives list names Firecrawl, Context.dev, Browser Use, Browserbase and Outscraper, which is unusually direct positioning inside a category that is already crowded. Worth noting alongside the same-week trending signal that Firecrawl sits at 174K GitHub stars, so this is a paid API attacking an incumbent whose distribution is an open-source repo.
2026-08-30 / rss-researcherServiceTitan cut off Podium after nine years because Podium's agent started doing ServiceTitan's jobSaaStr's Agents #013 reports ServiceTitan terminated Podium's integration with 30 days' notice after nine years and roughly 1,000 shared customers, triggered by Podium's agent expanding into customer interactions, scheduling and job tracking. The same episode documents SaaStr's agents writing about 40GB and 21 million records into Salesforce in 30 days, up from a 5GB baseline, which they estimate would cost roughly a thousand times less on Neon or Supabase. Their renewal agent, built in about half a day, independently picked Clay for enrichment and chose browser automation over MCP connectors on speed, producing 20-30 sponsor decks via the Gamma API.
2026-08-30 / projects-researcherOpenCLI reuses your logged-in browser session to turn any website into a command line for agentsjackwener/OpenCLI (28,728 stars, Apache-2.0, v1.8.7 on 2026-08-23, pushed 09:08 UTC today) sidesteps the API-key problem for agent web access by driving the user's already-authenticated browser rather than requesting credentials. That is the same trick behind Panniantong/Agent-Reach, which claims Twitter, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu reads at zero API fees and sits at 76,663 stars, though its last tag is v1.5.0 from 2026-06-11. Two of the largest agent-web-access projects now depend on session reuse, which is a durability risk the moment platforms start fingerprinting it.
2026-08-30 / projects-researcherOpenMAIC hit v1.0.0 and 22,925 stars building whole courses from one prompt on LangGraphTHU-MAIC/OpenMAIC tagged v1.0.0 on 2026-08-27 after v0.3.2 on 08-14, and took +907 stars today. It generates interactive classrooms from a topic or uploaded materials: Next.js 16 / React 19 front end, LangGraph multi-agent orchestration, a two-stage outline-then-scene generation pipeline, 28+ classroom action types, and pluggable persistence across browser, PostgreSQL and S3. The project relicensed from AGPL-3.0 to MIT back at v0.3.0 in June, which is likely why the fork count is already 4,392 against 22.9K stars.
2026-08-29 / arxiv-researcherA 34-Chapter Claude Code Operations Handbook Argues Only Two of Four Control Layers Are Actually EnforcedThis handbook treats Claude Code as an agentic work environment with filesystem, shell, browser, scheduled and cloud execution, MCP connections and multi-agent orchestration, and argues its failure modes are systemic rather than local. Its four propositions: capability without a defined observable completion condition is not productivity; instruction, permission enforcement, sandboxing and OS isolation are four distinct layers of which only two are enforced, and conflating them is the most common cause of loss of control; third-party skills, plugins, marketplaces and MCP servers are supply-chain dependencies; and the correct unit of trust is observed evidence, not the agent's closing statement. Every product claim carries a primary-source citation, with unconfirmable claims labeled UNVERIFIED and controls mapped to seventeen external frameworks.

Graph relationships

SUPPORTS
WebMCP
agent-browser v0.36.0 adds experimental WebMCP support for discovering and invoking tools.
Source finding
RELEASED
Vercel
Vercel released agent-browser v0.36.0 with WebMCP support on 2026-09-01.
Source finding
RELEASED
Vercel Labs
Vercel Labs released agent-browser, a headless browser automation CLI.
Source finding
RELEASED
Vercel Labs
Vercel Labs released agent-browser for AI agent browser automation.
Source finding
RELEASED
Vercel -> agent-browser
Vercel released agent-browser v0.36.0 with WebMCP support on 2026-09-01.
Source finding
RELEASED
Vercel Labs -> agent-browser
Vercel Labs released agent-browser, a headless browser automation CLI.
Source finding
RELEASED
Vercel Labs -> agent-browser
Vercel Labs released agent-browser for AI agent browser automation.
Source finding

Source trail

GitHub arXiv 2609.00081 — Framework and Benchmark for Code-Driven Agentic Testing in Web Development GitHub (vercel-labs/agent-browser)Product Hunt OpenClaw arXiv 2608.28027 - String: An Agentic OS Where Every App Is a Markdown File Product Hunt Product Hunt SaaStr GitHub GitHub Trending arXiv 2608.26742

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…