Give your AI agent a safer practice run
Before an agent touches your inbox, codebase or whole drive, give it a practice room. Today's three picks let you inspect a proposed tool call, make a UI's states visible, or search a deliberately small document collection. All have a no-paid-API first step on supported hardware. None makes an agent automatically safe. These are original-source-checked exercises, not our own model benchmarks or installation tests; keep real accounts, credentials and consequential actions out of the first run.
Try a tiny local agent model without giving it real tools
What changed
Tether AI Research published BotPsy on November 11: an Apache-2.0, 0.8B-parameter model tuned for function calling, based on Qwen3.5-0.8B. Public BF16 and quantized GGUF weights are available. Its documented examples return a proposed calendar tool call; the surrounding program still decides whether anything actually runs. That separation is the useful first experiment, not the vendor's leaderboard claim.
Why it matters
You can inspect an agent's choice and arguments locally before connecting it to a real service. A printed create_event call is not an event in a calendar. Keeping the executor absent lets you check a small model's behavior without granting email, file-write or payment access. It also makes a good place to test incomplete requests and notice when the model guesses.
Try it
With the Hugging Face CLI and a recent llama.cpp build supporting Qwen3.5 already installed, use a new scratch folder: hf download qvac/BotPsy-0.8B-GGUFs botpsy-0.8b-q8_0.gguf --local-dir . llama-server -m botpsy-0.8b-q8_0.gguf --jinja -c 32768 --temp 0.3 --top-p 0.95 --top-k 20 --presence-penalty 1.5 --chat-template-kwargs '{"enable_thinking": true}' --port 8080 In a second shell, ask for a fictional reminder, with no calendar connection: curl http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{"messages":[{"role":"user","content":"Add a practice reminder on Friday at 3 pm."}],"tools":[{"type":"function","function":{"name":"create_event","description":"Add an event to a pretend calendar","parameters":{"type":"object","properties":{"title":{"type":"string"},"day":{"type":"string"},"time":{"type":"string","description":"24h HH:MM"}},"required":["title","day","time"]}}}],"max_tokens":8192}' Read tool_calls and compare the arguments with your request. Do not implement create_event or expose this local server publicly. The command syntax above is for a POSIX-style shell.
Availability & cost: Public Apache-2.0 weights; no paid model API is required for local inference. The recommended Q8_0 download is about 0.81 GB, plus llama.cpp and additional working/context memory. CPU and supported GPU paths exist, but small download size is not a RAM or speed guarantee. First-time downloads need a connection.
Limits: Thinking must be explicitly enabled. The model card recommends English and low-to-medium-difficulty tasks, and excludes high-stakes automated decisions and sensitive uses. Its benchmark results are the developer's measurements, not ours. A well-formed call can still be wrong; we have not run this model or measured its latency on a laptop.
Original sources
- BotPsy: November 11 model announcement ↗Vendor announcement · Event 2026-11-11 · Checked 2026-11-11 09:11:11.705 UTC
- BotPsy model card: license, usage and limits ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:11.919 UTC
- GGUF downloads and the documented llama.cpp request ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:12.118 UTC
Give a coding agent your actual component states
What changed
Storybook announced version 11 beta on November 9. It adds official agent skills and headless CLI tools for finding component documentation, running tests and preparing browser reviews. The point is to let an agent reuse existing components and show what changed, rather than inventing another button from a screenshot. This is a beta with breaking changes, not a mandatory upgrade for a working app.
Why it matters
A UI can look convincing in its happy state while breaking when a label wraps, a field is empty or a button is disabled. Stories make those states explicit. The new docs commands give a coding agent a bounded source of component APIs; the same local workshop is useful for manual inspection without paying for an AI assistant or hosted visual-testing service.
Try it
Use a throwaway copy of an existing Storybook 10 project, not your production branch. Check the migration guide first, then follow the beta announcement: npx storybook@next upgrade After resolving any reported migration requirements, run: npx storybook@next tools docs list npx storybook@next tools docs show --id "your-component-id" Replace the quoted ID with one returned by docs list. Pick one existing button or input and inspect its normal, disabled and long-label stories in the local browser. Write down a missing state before asking an agent to change code. The CLI can also expose tests and review tools, but seeing docs or a preview is not itself a passing test.
Availability & cost: Storybook is MIT-licensed and can run locally without a paid AI model or Chromatic account. Version 11 requires Node.js 22.12+ and Vite 6.3+; the Vitest addon requires Vitest 4+. Framework requirements also apply. An optional coding assistant has its own access and cost; it is not needed for this first check.
Limits: Beta packages, addons and framework integrations can change. Older projects must reach Storybook 10 before this migration. The @next tag moves, so record the installed version and retain your lockfile. We inspected the announcement and migration guide, not a full upgrade. Test coverage only covers the states and assertions you actually provide.
Original sources
- Storybook 11 beta: November 9 announcement ↗Project announcement · Event 2026-11-09 · Checked 2026-11-11 09:11:12.428 UTC
- Storybook 11: breaking changes and prerequisites ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:12.548 UTC
- Storybook source: MIT license ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:13.270 UTC
Find a file by its contents, without indexing your whole life
What changed
Kensaku's November 9 v1.0.17 security release hardens agent scope and direct reads, update staging, watched-file exclusions and Office extraction. The MIT-licensed desktop launcher searches filenames, document text and screenshot text locally, with optional semantic search and MCP access. The security release is the reason to inspect the current version rather than installing an older copy from a roundup.
Why it matters
A useful agent should retrieve the right note without needing your entire home directory. Local content search can help you rediscover a setup instruction or error screenshot, and a deliberately tiny corpus makes wrong matches easier to spot. Local storage is not automatic confidentiality: the index contains extracted text, and anything a connected agent reads may become context for that agent's model provider.
Try it
Read the v1.0.17 release and security notes before installing. If your device policy permits the unsigned build, obtain the matching release asset and verify it against the attached SHA256SUMS; do not disable OS protection to get it running. Create one disposable folder with two invented notes and a screenshot containing a made-up error phrase. During setup select only that folder. Search for a phrase inside a note, then the text in the screenshot; inspect the highlighted snippet in the actual file. Leave MCP and the optional local LLM disconnected for this first pass. Do not select your home folder, browser profile, work drive or credential store.
Availability & cost: Free MIT software, with release assets for Windows, macOS and Linux. Keyword/content search needs no paid model API or account. Semantic/OCR features may download local models and use extra disk and RAM. Windows source builds require C++ build tools and WebView2; check the install guide for other platforms.
Limits: Release binaries remain unsigned; if your system blocks them, stop or use your organization's approved source-build process. The index itself is sensitive. Exclusion rules and read-only MCP are boundaries to inspect, not a promise against every leak. Windows gets more real-world testing than macOS/Linux according to the maintainer. We have not installed or benchmarked Kensaku.
Original sources
- Kensaku v1.0.17: November 9 security release ↗Project announcement · Event 2026-11-09 · Checked 2026-11-11 09:11:13.873 UTC
- Kensaku: local search, platforms and setup ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:14.869 UTC
- Kensaku threat model and sensitive index data ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:14.297 UTC
4 more signals
Useful checks and updates. Guidance shows its check date; event dates identify actual announcements.
Make an agent's evidence chain visible
Agent Cockpit is an AGPL-3.0-or-later local whiteboard and MCP server. Its demo records runs and claims, then shows which claims become stale when dependent code changes. Start with the supplied demo, not your main repo: init writes agent configuration and Git hooks. Node.js 22.13+, Git and Python are needed for the demo; its deterministic checks do not require model tokens. A traceable claim is still not proof the experiment was well designed.
- Agent Cockpit: demo, prerequisites and license ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:15.372 UTC
Record a tiny UI bug instead of describing it badly
The Apache-2.0 chrome-capture project records a Chrome tab to AV1 MP4 using the browser's recording API. Its documented first command is: npx chrome-capture https://paulirish.github.io/chrome-capture/ demo.mp4 4s. It needs Chrome 153+ and Node.js 24.11+. Use a public demo, not a logged-in dashboard; audio is on by default. High frame-rate claims depend on hardware, and Windows performance is not established in the README.
- chrome-capture: public demo command and platform caveats ↗Vendor guidance · Current guidance · Checked 2026-11-11 09:11:15.954 UTC
Find what blocks an AI-written pull request
GitHub's November 5 opt-in Overview preview moves merge readiness, checks and approvals to the top of a pull request. If the preview is available to your account, use Try the Overview page and inspect one blocking check or unresolved conversation. You can return to the classic view. This is a review-navigation change, not a substitute for reading the diff or a reason to merge automatically; available actions still depend on your permissions.
- GitHub: new pull request Overview public preview ↗Vendor announcement · Event 2026-11-05 · Checked 2026-11-11 09:11:16.096 UTC
Turn one paper sketch into a throwaway prototype
Google's November 5 walkthrough uses a photo of a sketch, Canvas and the prompt Can you code this app? Draw a three-box habit tracker with no personal data, upload it, then test every control in the preview. Canvas has access for free Gemini users, subject to account, region and usage limits; paid plans offer different model/context access. Do not add logins, payments or real user data to this first experiment, and do not treat generated code as production-ready.
- Google: sketch-to-app Canvas walkthrough ↗Vendor guidance · Event 2026-11-05 · Checked 2026-11-11 09:11:16.416 UTC
Build something cool
Write a one-task practice contract
Use the free agent-handoff workflow to specify one of today's experiments: a fictional reminder call, a component-state inspection or a search over two fake notes. Record the allowed input, forbidden actions, expected observable output and one failure case. Run only the chosen first step, compare what actually happened with that contract, and keep the result for a later rerun. No real inbox, payment tool or whole-drive access belongs in this practice run.
Open the DistroCool workflow →Was this edition useful?
Your choice helps us shape future editions. No account needed.