Agent Snacks — Snack Pack #1
🍪 Week of Oct 9–16, 2026 · 7 snacks · ~4 min
Agents learned to babysit themselves this week. Grab a coffee — they’ll keep chewing.
Hey — welcome to Snack Pack #1. Same deal every Thursday: only what actually shipped in the AI coding-agent tools you already use, each bite with a source link. No funding fluff. No “AI will change everything.” Just the version bumps that might change your Friday.
Nutrition Facts
Serving: 1 Snack Pack · ~8 min
Counts from this issue. “Actually useful” is the editor's call: 5 of 7 change a workflow. Hype: 0 g.
- nice to know
- worth a look
- act this week
First Bite 🍪
Use Cline and never picked a model? Surprise: since v4.1.24 your default is now Claude Sonnet 5.5 (release notes). Same kitchen, new chef, so answers and costs can taste different. Try this: open your AI tool's settings and pick the model yourself. Thirty seconds, no surprises. New to all this? Start here.
1. Copilot will now farm your CodeQL backlog for you (GA, Oct 15) 🌶️🌶️
Real workflow shift, but enterprise-policy gated
What changed: Assign a CodeQL alert to Copilot and it wanders the repo, proposes a fix, re-runs CodeQL to prove the alert dies, then opens a draft PR. Batch up to 25 alerts into one PR, or poke it via the REST API (copilot-swe-agent[bot]). Default-branch alerts only; an enterprise admin has to flip on the “Agentic autofix” policy; it spends Code Security AI Credits + Actions minutes.
Who should care: Security folks staring at a CodeQL mountain on GHEC / GHAS.
Source: GitHub Changelog
2. HydraFusion stops making you pick a model (and adds a “Max” mode) (Oct 14) 🌶️
Research preview; nice, not urgent
What changed: Copilot’s multi-model orchestrator (still research preview) refreshed its pool — Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna — drafts on medium effort, reviews on high. Cascade gets a judging step; Critique can tool-snoop the workspace with parallel critics. New HydraFusion Max is the “go deep” button. Lives in VS Code, the Copilot app, and Copilot CLI.
Who should care: Copilot Pro+ / Business folks tired of the model roulette wheel.
Source: GitHub Changelog
3. Claude Code: idle chats now compact before the cache expires (v2.1.299–300, Oct 14) 🌶️🌶️
Automatic on long sessions; know the opt-out
What changed: Leave a 200K+ token terminal chat alone ~54 minutes and it compact-snacks itself before the one-hour prompt cache dies — so your next message doesn’t re-chew the whole transcript. Tune with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS or kill it with "idleCompaction": false. Also: CLAUDE_CODE_DISABLE_FILE_MENTIONS (paths only, not file guts), and OTel now splits cache writes into cache_creation_5m_tokens / cache_creation_1h_tokens.
Who should care: Anyone with marathon Claude Code sessions — or a token bill that makes them wince.
4. Copilot CLI: hooks that don’t hang the babysitter (v1.0.96–98, Oct 12–14) 🌶️🌶️🌶️
Unattended Autopilot loops stop hanging
What changed:
- v1.0.96: /add-dir grants sandbox access for the session; interactive settings nudge you to mask likely secrets.
- v1.0.97: agent-scoped declarative hooks; Autopilot now denies a preToolUse “ask” instead of freezing on a dialog nobody is watching. 🫡
- v1.0.98: Gemini 4 Argon support; /ask expands with Ctrl+E, copies with Ctrl+Y.
Who should care: Terminal Copilot users running Autopilot / unattended loops.
Source: v1.0.96 · v1.0.97 · v1.0.98
5. Cursor Origin opens the snack aisle (App Marketplace, Oct 13) 🌶️
Matters mostly to tool makers and CI wiring
What changed: Origin plugs into 20+ apps — Cloudflare, Vercel, Buildkite, CircleCI, CodeRabbit, Greptile, Semgrep, Socket, Sentry, Linear, and friends. Public submissions welcome. Apps auth with app JWT + install tokens, get webhooks, write check runs / PR comments / reviews.
Who should care: Dev-tool makers hunting distribution, and teams wiring CI/review bots into Origin.
Source: Cursor changelog
6. Cline quietly swapped your default model (v4.1.24, Oct 13) 🌶️🌶️🌶️
Unpinned users are silently on a new model
What changed: No model pinned? Cline + OpenRouter now default to Claude Sonnet 5.5. Catalog refresh adds Haiku 5.5; ~20 other providers also shuffle defaults — unpinned users will wake up on a different model. Bonus: turning reasoning off no longer 400s on several frontier models. Desktop v0.0.46 adds a clickable prompt outline.
Who should care: Cline users who never pinned a model. Check the cost dashboard before you blame the agent.
Source: v4.1.24 · desktop-v0.0.46
7. OpenCode 1.19: the on-ramp to OpenCode 2 (and auto-update is off) (Oct 11) 🌶️🌶️
Upgrades are now manual; plan the jump to 2
What changed: opencode upgrade --major (or /update in the TUI) walks you to OpenCode 2. Updates are no longer automatic — upgrade on purpose. Also allows OpenAI’s ultrafast service tier.
Who should care: OpenCode folks. Deliberate upgrades > surprise majors.
Source: v1.19.0
The Main Course 🍽️
MCP vs plugins vs skills: what actually plugs your agent into the world?
Jargon jar: an agent is an AI that does things, not just chats. A tool is one thing it can do (like "search issues"). The context window is everything it can see at once: its short-term memory, or plate size (Anthropic docs).
Every tool on the shelf this week claims to "connect your agent." They connect it in three different ways, and picking the wrong one is how you end up with 40 tools in context and an agent that still can't find your deploy script. Here's the menu.
MCP servers = the hands. The Model Context Protocol is an open standard for connecting AI apps to external systems: data, tools and workflows. The official docs call it a USB-C port for AI, and Claude, ChatGPT, VS Code and Cursor all support it. A server exposes tools (functions the model can call), resources (data it can read) and prompts (templated workflows). In Cursor that means a stdio command for local servers, or a Streamable HTTP/SSE URL for remote ones, set in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global) (Cursor docs). Write once, plug into any MCP client. That's the whole pitch.
Use MCP when the agent needs live data or needs to do something: open a PR, query a database, read a ticket. Make it a remote HTTP server when a team will share it.
Skills = the recipe card. An Agent Skill is a folder with a SKILL.md, which has YAML frontmatter (name, description) plus instructions, and can bundle scripts and reference files. The trick is progressive disclosure. Only the name and description load at startup (~100 tokens per skill). The body (under 5k tokens) loads when a request matches. Bundled files cost nothing until they're read, and scripts run through bash so only their output enters context. In Claude Code, skills live in ~/.claude/skills/ (personal) or .claude/skills/ (project).
Use a skill when the problem is know-how, not access: "how we write migrations," "our release checklist," "fill this PDF form." Skills teach the agent the moves. They don't give it new hands.
App plugins/connectors = the meal kit. Plugins bundle the other two for one app. OpenAI's plugin docs describe ChatGPT plugins as skills + MCP servers + optional UI, packaged and submitted to the ChatGPT Directory, with a guide for converting a Claude Code plugin. Cursor's Marketplace offers official plugins with one-click install and OAuth, and team admins can distribute MCP servers through a team marketplace (Cursor docs).
Use a plugin when you want distribution and polish inside one app. Underneath, it's still MCP and skills, so if you need portability, ship the MCP server too.
The cheat sheet
| Need | Reach for |
|---|---|
| Live data / take actions | MCP server |
| Repeatable workflow or house rules | Skill |
| One-click install inside one app | Plugin / connector |
| Works across Claude, ChatGPT, Cursor | MCP server (+ skill) |
Gotchas that bite - Context bloat: every MCP tool you enable is another option the agent has to weigh. Cursor lets you toggle servers off without removing them, and enterprise admins can allowlist servers and specific tools. - Skills don't sync: custom skills on claude.ai, the Claude API and Claude Code are managed separately. Upload once per surface. - Sandbox differences: skills on the Claude API run with no network access and no runtime package installs. In Claude Code they get your machine's network. - Trust: both Anthropic and Cursor say the same thing: only install skills and servers from sources you trust, and audit the code. A skill can run scripts. An MCP server can act on your behalf. Treat both like installing software.
Chef's pick: start with one remote MCP server for the system you touch most (GitHub, your DB, your tracker), then write one skill that says how your team uses it. Hands plus recipe beats a drawer full of gadgets. 🍽️
Secret Sauce 🥫
Try it now — Claude Code idle compaction (v2.1.299+):
1. Update Claude Code to v2.1.299 or later.
2. Idle 200K+ token terminal chats now compact after ~54 minutes, before the 1-hour prompt cache expires — so step away guilt-free.
3. Want the full transcript kept? Set "idleCompaction": false, or raise the threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS.
Video Tasting 📺
Claude Agent Skills Explained · 🟡 Intermediate · 3:14 · Anthropic (official)
After watching you can: Tell skills, CLAUDE.md, MCP servers and subagents apart, and know when to use each.
- Skills are folders of expertise Claude loads automatically when a task matches.
- Only each skill's name and description load at startup; the full SKILL.md loads on demand.
- Compares skills with CLAUDE.md (project facts), MCP (connects to data) and subagents (fixed roles).
Short on time? Watch just: 0:30 How Agent Skills work · 1:42 Agent Skills vs MCP Servers · 2:33 Stacking skills into one bigger agent
On screen: mostly a code editor: skill files, CLAUDE.md, an mcp.json with Playwright, and skills stacked into one bigger agent.
Summary · Anthropic. More videos: /videos/
New on the Shelf 🛒
Fresh MCP servers, skills, plugins and templates, each checked against its primary source.
- MCP server GitHub MCP Server v2.0.0 🌶️🌶️ ✓ verified
GitHub's official server now ships structured outputs and output schemas, so Code Mode agents get typed results instead of guessing (schemas need MCP 2026-07-28+ clients). - MCP server MCP reference servers 1.0 🌶️ ✓ verified
The classics grew up: filesystem, memory, everything and sequential-thinking hit 1.0.0 on npm; fetch, git and time got fresh PyPI builds. - Skill Anthropic claude-api skill 🌶️ ✓ verified
Claude's cheat sheet for building on Claude learned dynamic workflows, four workflow quickstarts, and that Haiku 5.5 is the current Haiku.
Quieter on the tray: Gemini CLI v0.64.0 tightens folder-trust in headless mode (releases). Devin Desktop 3.10.67 speaks a newer MCP revision and lets admins pin the extension marketplace. Codex CLI shipped 0.162.1 bugfix + a pile of 0.163.0 alphas. No Aider release in the window.
🍪 Pattern of the week: The agents got a babysitter. Auto-denied permission prompts, idle compaction, self-verifying autofixes — this week was less “smarter model,” more “you can actually walk away.”
That’s the box. Forward a snack to a teammate who lives in the terminal.
Unsubscribe anytime via the link in this email (beehiiv) — we keep the list tidy on purpose.