Check a skill. Build a UI. Test one tab.
A useful agent experiment ends with something you can inspect, not a bigger permission list. Start by checking whether a skill can load, build a tiny search field with a clear keyboard behavior, then give a browser agent one harmless page. The software is free; installation, compatible runtimes and your chosen model still have requirements. Keep the first attempt in scratch files and a browser profile with no accounts signed in. We inspected original release notes and documentation, not independently measured these tools or their performance.
Find out whether an agent can load your skill
What changed
NVIDIA published SkillEvaluator v1.0.1 on November 18, its first complete release in the 1.0 series. HarnessCheck compares skill locations, frontmatter, descriptions and content limits against versioned agent profiles. It can flag a skill that a particular harness may miss, reject or truncate. This is a compatibility check, not a model judging whether the instructions are good.
Why it matters
A carefully written skill cannot help if the agent never sees it. Checking the file contract first gives a builder a concrete place to start before spending tokens on a full evaluation. Keep two questions separate: whether the harness delivers the skill, and whether following it actually improves a task.
Try it
In a separate Python 3.12–3.14 virtual environment, review the package and install the documented CLI; pinning this edition uses: python -m pip install skillevaluator==1.0.1 Copy a small skill you wrote into a scratch folder named my-skill. Do not use private customer material. From its parent directory, run: skillevaluator harness-check ./my-skill Read the result for your actual agent and its supported version. Fix one reported location or frontmatter problem in the scratch copy, then rerun and compare the finding. Optionally run: skillevaluator quality-check ./my-skill Keep any generated report outside my-skill. Stop at these static checks; do not enable all tiers, provision a model, install a skill into a real agent or connect accounts for this first attempt.
Availability & cost: Apache-2.0 source; version 1.0.1 is published on PyPI for Python >=3.12 and <3.15. HarnessCheck needs no model key or agent run. The lightweight quality check needs no API key, security scanner, GPU or Docker. Initial package installation requires downloads; these checks do not require paid inference.
Limits: Harness profiles describe particular releases, not every future agent version. A compatibility result is not evidence that the skill was invoked successfully. A quality score is not a security audit or a measured improvement: publication evidence remains INCOMPLETE until additional checks run. Full validation has separate scanner, model and sandbox requirements. We did not install or run this CLI.
Original sources
- SkillEvaluator v1.0.1 release ↗Project announcement · Event 2026-11-18 · Checked 2026-11-18 09:09:39.831 UTC
- Pinned 1.0.1 README: compatibility, Python and keyless checks ↗Official guidance · Current guidance · Checked 2026-11-18 09:09:40.526 UTC
- Pinned keyless quickstart and report limitations ↗Official guidance · Current guidance · Checked 2026-11-18 09:09:40.915 UTC
Make a Clear button that gives focus back
What changed
Vercel Labs has released UI SDK as a public beta. Its core package first published version 0.1.0-beta.1 on November 12. Instead of only offering finished components, the library exposes behaviors for focus, selection, popups and keyboard interaction, with framework adapters. The official search-field example combines an input with a conditional Clear button and focus recovery.
Why it matters
A tiny interface is a better first check than asking an agent to rewrite your app. Clearing text is simple to describe, but the keyboard user also needs focus to return somewhere useful. You can inspect and test that behavior before adopting a component system or allowing an agent to operate the interface.
Try it
Use a disposable React 19+ project, not your production app. After reviewing the beta and your dependency policy, install the two documented packages: npm install @ui-sdk/react @ui-sdk/core Use the complete search-field and React examples in the linked README, or give an already available free/local coding assistant this original bounded prompt: In this scratch React project, use UI SDK to make a labeled Search input and a Clear button. Show Clear only when text exists. Clearing must empty the input and return focus to it. Follow the installed package documentation and keep the existing styles. Do not add analytics, accounts, an agent bridge or a deployment. Start the project with its existing development command. Type a fictional query, reach Clear with Tab, activate it with the keyboard, and confirm that the input is empty and focused. Also try a pointer click and an initially empty field. If dependency-age rules block installation, stop rather than disabling them.
Availability & cost: Apache-2.0 library with React, Vue, Svelte, Solid and Web Components adapters. The React quickstart requires React 19 or newer and a compatible project toolchain. The library needs neither paid hosting nor a model API. Any coding assistant is optional and has its own access terms; the README example can be copied manually.
Limits: The beta APIs, documentation and behavior can change, and the project also points to Vercel beta terms. Built-in behaviors do not establish that your composition is accessible. Test labels, keyboard use and focus in the actual app. We inspected the examples but did not compile this component. No claimed reduction in code, tool calls or cost is our own measurement.
Original sources
- npm registry: core beta publication timestamp and license ↗Project announcement · Event 2026-11-12 · Checked 2026-11-18 09:09:42.142 UTC
- UI SDK README: complete search-field example and beta terms ↗Official guidance · Current guidance · Checked 2026-11-18 09:09:42.491 UTC
- UI SDK quickstart: React requirements and component prompt ↗Official guidance · Current guidance · Checked 2026-11-18 09:09:42.893 UTC
Give OpenCode Browse one clean tab, not your accounts
What changed
OpenCode Browse v0.1.0 was released on November 16 with a Rust rewrite of its CLI, local relay and MCP server. A Chromium extension still drives the browser. Tabs replace the earlier session handle: agents work with the tabs they open and tabs a person explicitly lends them. The release describes a smaller native runtime; that is not an independently verified speed or memory result.
Why it matters
An agent using your ordinary browser can inherit access to real logins and actions. A dedicated test profile lets you inspect the new browsing path without handing over your inbox or payments. Tab ownership helps organize work, but the safer first boundary is having no valuable signed-in session available at all.
Try it
First create a separate Chromium-browser test profile with no accounts signed in. Review the official extension permissions and installer changes. The project documents: npx @opencode/browse install Choose only the intended test agent in its setup picker and pair the extension in that empty profile, not your everyday browser. Setup can edit agent configuration and PATH; inspect those choices before accepting them. Use an MCP-capable agent you already run without paid model calls. Give it this original task: Open https://example.com in a new tab. Read the page title and heading, then report the tab you used. Do not sign in, submit forms, download files, buy anything or inspect other tabs. Stop after the report. Compare its answer with the visible page yourself. Do not lend it existing tabs. If it needs an account or a paid model to proceed, stop rather than expanding this experiment.
Availability & cost: MIT-licensed bridge and a Chromium extension; documented browsers include Chrome, Edge, Brave and Chromium. The native runtime does not use Node.js or Playwright underneath, but the npm launcher requires Node 22+. It does not include a model: use an existing free/local MCP-capable setup, with its separate hardware and model terms.
Limits: The extension requests powerful debugger and website permissions. Material read from a page can reach the agent model provider, and page text can contain prompt injection. A separate profile and a prompt are not a complete sandbox. Installation and automatic updates change local software/configuration; review them. We did not install Browse, verify its isolation or reproduce the headline memory claims.
Original sources
- OpenCode Browse v0.1.0 release ↗Project announcement · Event 2026-11-16 · Checked 2026-11-18 09:09:41.544 UTC
- Pinned Browse README: install, tabs, permissions and model boundary ↗Official guidance · Current guidance · Checked 2026-11-18 09:09:41.894 UTC
4 more signals
Useful checks and updates. Guidance shows its check date; event dates identify actual announcements.
Kitesurf is open source, not a free hosted browser
Cloudflare open-sourced its experimental browser engine on November 13 under Apache-2.0. Local development uses Wrangler; the README requires Bazelisk, Node 22+ and pinned pnpm, with a substantial first build. Read those requirements before trying the local health endpoint. Deploying it on Cloudflare needs a Workers Paid plan because of Dynamic Workers. CDP support is not full Chromium compatibility; do not substitute it for your only real-browser QA.
- Cloudflare: Kitesurf open-source release and local/paid boundary ↗Vendor announcement · Event 2026-11-13 · Checked 2026-11-18 09:09:43.099 UTC
Turn a small source set into an interactive study page
Google announced Gemini Notebook sites on November 16. On the web, an eligible user aged 18+ can ask the notebook chat to make a site from sources they can edit. Try a fictional project checklist and inspect every generated date. Standard access exists without a paid AI plan, but rollout and compute quotas apply; stop when the free allowance is unavailable. Sharing a site requires notebook access, so never publish one containing private sources.
- Google: source-based interactive sites rollout ↗Vendor announcement · Event 2026-11-16 · Checked 2026-11-18 09:09:43.918 UTC
Agent identity is not account permission
Poppy 1.0 separates session identity from OAuth account access. Read its discovery and authorization sections, then write a local table for one toy app: what is public, what needs read access, and what changes need explicit approval. Reading the public specification costs nothing. A recognized agent or a valid session does not itself authorize transactions, and a protocol description is not proof your implementation is secure.
- Poppy 1.0 specification: discovery, sessions and authorization ↗Official guidance · Current guidance · Checked 2026-11-18 09:09:44.363 UTC
Plan for a malicious page before opening it
Microsoft recommends layered defenses for indirect prompt injection, including isolating untrusted text, least privilege, tool-chain checks and human review of risky actions. For your next browser test, write down the forbidden actions and keep sending, purchasing and account changes unavailable. The guidance is free to read; the named enterprise services are not required for that checklist. No prompt or scanner guarantees that every injection is stopped.
- Microsoft: layered indirect prompt-injection defenses ↗Vendor guidance · Current guidance · Checked 2026-11-18 09:09:44.820 UTC
Build something cool
Make one tiny interface pass a real user journey
Use the linked free website-check workflow for your scratch search field. Specify three checks: an empty field has no Clear control; typing shows Clear; keyboard activation clears the query and restores input focus. Inspect the actual browser state and record failures, not just an agent report. Keep any browsing agent in a separate signed-out profile with no form submission or external actions authorized. This is a small acceptance checklist, not a security certification or a production release.
Open the DistroCool workflow →Was this edition useful?
Your choice helps us shape future editions. No account needed.