10 JC Lab Ideas to Film, Test and Write Today: Coding Agents, Agent Guardrails and Safer Trend Checks

Video topics to film

1. fx on a small machine — How-I-tested

• Subject: Vercel's experimental fx, a Zig-based, open, native coding-agent harness.
  • Film today: Install it in a disposable project, run the same three prompts through fx ask, then repeat one task in an ordinary terminal agent. Show cold start, binary size, first useful edit, tool approvals, and the number of manual corrections.
  • Hook and verdict frame: “Does a 7.8 MiB native agent feel faster, or does it simply move complexity into setup?” The official repository describes fx as model-agnostic, usable with local or cloud inference, and experimental; its README also shows Vercel AI Gateway and Codex login paths. Treat those as capabilities to verify, not proof of quality. 1 The project also drew a large Hacker News discussion, which is useful discovery context rather than a benchmark. 2

2. OneCLI: can an agent work without seeing the key? — Tutorial

• Subject: OneCLI's sandboxed, gateway-based agent setup for teams.
  • Film today: Self-host the open-source project if the machine can run its prerequisites, create one harmless test agent, grant access to a throwaway repository, and attempt three actions: read a file, call an approved service, and request a blocked action. Show the approval prompt, gateway decision, audit trail, and what appears in the agent's context.
  • Hook and verdict frame: “The important screen is the credential boundary, not the chatbot reply.” OneCLI's README says the real credential is injected at the gateway per request and kept out of the agent's context, memory, and logs; it also documents sandboxing, human approvals, and self-hosting. 3 The founders described the same design and its prompt-injection motivation in their Launch HN post. 4 A demo with real Gmail, cloud, or exchange keys would prove too much at once, so use throwaway credentials and a fake target.

3. ChatGPT controlling iMessage on a Mac — Scam check

• Subject: A privacy-and-permissions check for the reported ChatGPT–Apple Messages integration.
  • Film today: Use a spare macOS profile with two dummy conversations. Before enabling anything, record every consent screen. Then test reading, searching, summarising, drafting, and sending a message; revoke Full Disk Access and the related automation permission one at a time, and show which action stops working.
  • Hook and verdict frame: “What exactly does the integration need before it can touch a message?” Bloomberg reported that the Mac feature can read, write, send, search, and summarise Messages, while requiring consent plus permissions for message history, Full Disk Access, contacts, and automation tools. 5 The video should verify the permission path on camera. It should avoid real conversations and avoid presenting a news report as proof that every account has the same rollout.

4. Google Antigravity leaves its own desktop app — Trend react

• Subject: Google's newly expanded Antigravity workflow inside familiar IDEs.
  • Film today: Open the same small repository in VS Code and, if available, one preview editor. Ask the agent to make one documented change, inspect the permissions and audit settings, and compare the workflow with the standalone app or CLI.
  • Hook and verdict frame: “Does bringing the agent into the editor remove friction, or remove a useful safety boundary?” Google Cloud's August 20 announcement says the new extensions cover VS Code, with Visual Studio, JetBrains, and Zed listed as previews, alongside the desktop app and CLI. The announcement also describes workspace sandboxing, MCP and browser controls, audit logging, and spend controls for eligible Gemini Enterprise subscriptions. 6 Separate what you can try on an individual account from enterprise-only controls.

Tools and apps to review

5. Router by Ramp — Review

• Subject: A new model-routing service positioned as infrastructure for choosing among AI models.
  • Film today: Use one fixed prompt set with a direct provider call and the Router path. Record the selected model, response time, input and output tokens, retry behavior, visible price, and whether the route can be pinned. Use the same temperature and output limit where the interfaces allow it.
  • Hook and verdict frame: “A lower invoice is useful only if the route remains observable.” Router was a fresh Hacker News lead when this pack was prepared, with the submission linking to router.com. 7 Review the service's own terms and pricing before recording a cost conclusion. A launch discussion is a discovery signal; it is not independent evidence that routing saves a fixed percentage.

6. OSV-Scanner for an AI project's dependency tree — Comparison

• Subject: Google’s open-source vulnerability scanner, compared across a source directory and a container image.
  • Film today: Run osv-scanner scan source -r against a small JavaScript or Python project, then scan the image that runs it. Capture package name, advisory, affected range, fixed version, exit code, and the time needed to produce the report. Add one deliberately old test dependency in a disposable branch.
  • Hook and verdict frame: “Which check catches the problem you actually ship: the lockfile scan or the image scan?” The official repository says OSV-Scanner connects dependency lists to the OSV database, supports many language ecosystems and package managers, and can scan source directories and container images. It also marks guided remediation as experimental and warns that fixes can run package-manager scripts on untrusted projects. 8 Compare detection scope and remediation risk, rather than declaring one command the universal security answer.

7. PostHog as an AI-workflow observability layer — Review

• Subject: PostHog's analytics, session context, AI observability, and MCP-facing workflow.
  • Film today: Instrument one tiny AI web app with three events: prompt sent, tool call approved, and result accepted. Compare the dashboard trace with raw application logs, then inspect what an MCP-connected workflow can read or change. Keep the test dataset synthetic.
  • Hook and verdict frame: “Can a creator explain why an AI feature failed without opening five different dashboards?” The official repository describes AI observability, analytics, session replay, error tracking, logs, and steering from Slack, web, desktop, or MCP. 9 Treat that description as the product surface to test. The review should show setup effort, trace completeness, and data-retention choices instead of repeating a feature list.

Trending niche subjects

8. Aligned and Pump.fun: attention versus a tradable setup — Trend react

• Subject: Two names surfaced by the current crypto trend pull, handled as a paper-trading comparison rather than a buy call.
  • Film today: Open the live pages for Aligned (ALIGN) and Pump.fun (PUMP). Record the same fields for both: price, 24-hour volume, market-cap rank, spread or liquidity warning, seven-day range, and the exact timestamp of the screenshot. Define one entry, one invalidation level, and one exit rule on paper before looking at the next candle.
  • Hook and verdict frame: “Can a popular search lead survive a basic liquidity and volatility check?” CoinGecko provides live market-data pages for Aligned and Pump.fun, including chart, market-cap, and volume fields. 1011 Use the trend as a queue for research. Search popularity is not evidence of a safe entry, future return, or adequate execution liquidity.

9. Argentic and the pay-per-request web — How-I-tested

• Subject: A Show HN project described as an L402 Lightning toll booth for AI scraping agents.
  • Film today: Test one public page or a local mock endpoint. Show the request before payment, the HTTP 402 response, the invoice or payment instruction, and the response after a test payment only if the amount and network are fully understood. A no-payment walkthrough is enough for the first cut.
  • Hook and verdict frame: “Can a creator charge an agent for one useful response without building a full subscription system?” The Show HN page identifies Argentic as an L402 Lightning toll booth for AI scraping agents. 12 The post is a builder announcement with very early discussion, so the film should test the request flow and explain fees, wallet custody, refunds, and abuse controls. A working 402 response proves a payment gate, not recurring income.

Article to write

10. How to test an AI model that returns a successful blank answer — How-I-tested

• Subject: A reproducible guide to detecting zero-visible-byte responses without confusing them with refusals, safety blocks, or network errors.
  • Write today: Define “blank” as a successful response containing zero visible UTF-8 output bytes. Use the same prompt in a small matrix of models, log status, finish metadata, raw byte count, retries, and provider name, and keep the raw records. Add a control prompt that explicitly asks for a short visible answer. Publish the test script and a table that separates blank responses from refusals, tool runs, and operational failures.
  • Reader promise: “You will know whether a model is actually silent, whether your parser hid the answer, and which evidence belongs in a bug report.” A Zenodo preprint published July 29, 2026 reports 11,658 zero-visible-byte outcomes across 31,430 completed trials and explicitly separates them from refusals, safety blocks, tool-mediated executions, and infrastructure outcomes. 13 Use the paper as a reproducibility lead, not as proof of an internal model mechanism. The article should report the observed response states and stop before assigning a cause.

Notes for the edit

Film in this order

1. Capture the fx install and first prompt.
2. Record the OneCLI permission boundary with throwaway data.
3. Run the macOS iMessage permissions check.
4. Shoot the Antigravity editor comparison while the setup is fresh.
5. Batch the OSV-Scanner, Router, and PostHog screen recordings.
6. Finish with the paper-trading comparison and the Argentic payment-gate walkthrough.
7. Write the blank-response article from the logs collected during the day.