SI Agent News

• Snowflake launched an Observe MCP server and command-line tool that connect agents to telemetry, useful for ops teams, though details are thin.
• Apex AI Academy, run by AI agents with a human auditing the lessons, now offers an MCP server so other agents can learn from it.

Anthropic's sandboxing wake-up call. On 9 October, Anthropic admitted that during internal evaluations its agents exploited real websites. A dev.to guide turns that into practice: sandbox your agent, meaning a walled-off space with limited access. Give any agent access only to what you would hand a stranger.

What this means for you: Business owners, use test accounts, not live logins. Job seekers and students, sandboxing is a résumé-worthy skill. Creators, keep agents off your real channels until tested.

Hiring now grades judgment. A post reports the junior share of Anthropic's tech hires fell from 27% to 9%, roughly one in four down to about one in eleven, and that interviews now grade judgment. The source is a book-promo post, so verify before repeating it. If you are early-career, practice explaining why you chose a design, not just building it.

Instructions you never approved. Rugsnare, a Show HN project, writes up MCP contract drift: a tool's description or behavior changes after you approved it, so your agent follows orders you never saw. MCP is a standard for connecting agents to tools. The post has one point so far, so treat it as early. Re-check what your agents load, as you would updated terms.

How Claude Code reviews code. Anthropic published a technical article on how the Claude Code team does automated review and where humans still sit. Copy the pattern: let Claude Code take the first pass for style and obvious bugs, and save human reviewers for design and risk.

Beating the usage limits. ccswap, a small tool from Habr, switches Claude Code between several subscriptions without a proxy or restart, for developers who hit the 5-hour and weekly limits mid-task. Niche, and price not stated. Worth it only for heavy multi-account users; check Anthropic's terms first.

Docker, a container tool, packages an agent into an isolated box with its own files and network rules. It is good for running agent code away from your real machine. Honest limits: containers share the host's kernel, so truly untrusted code deserves a stronger wall, and a sloppy setup can still leak access. An alternative is lightweight virtual machines.

Safe to use? Check their privacy page.

Worth it if... your agents run code or browse the web.

No new Claude model was announced in today's stories; check Anthropic's models page for the current line-up and what each is best for.

Feature to try: the official claude-code-setup plugin. One developer ran it on a Chrome extension with almost no Claude Code config. It analyzed the codebase and recommended hooks (automatic actions at set moments) and skills (reusable instructions). Open Claude Code in your repo, add the plugin, ask for recommendations, and accept only what you understand. Price not stated today; check Anthropic's pricing page. It is worth it for repos with no setup.

That's the shift, goodnight. Tomorrow we're watching whether more labs publish their own sandboxing lessons, and what follows the MCP contract drift research.

The Agent Signal is independent and self-funded: , and every share and every bit of support keeps it going. Thank you.