$ cat ./issues/005

Nov 12, 2026

Issue 005: Fable got cheaper, Gemini 4 Carbon lands

buildwith.sh // issue 005 // Nov 12, 2026

30 new drops since last Thursday. Anthropic's Fable 5.2 got cheaper, Google answered with Gemini 4 Carbon, and Cursor now lets a whole team share one fleet of agents. Here are the ten worth your time.

$ ls ./new --since=last-issue --top=10

Claude Fable 5.2 by Anthropic [Nov 9 · agent]
Anthropic's new Fable model for hard coding and long-horizon agent work went GA on Nov 9 as claude-fable-5-2: 1M-token context, 128K max output, always-on adaptive thinking, and outputs up to 30% faster than Fable 5.1. It's cheaper at $8/$40 per million input/output tokens (Fable 5.1 was $10/$50), with an optional fast mode at $16/$80. It's on the Claude API, Bedrock, Google Cloud and Microsoft Foundry, and GitHub Copilot added it the same day for Pro+, Max, Business and Enterprise. It requires 30-day data retention, and Mythos 5.2, the same model with fewer safeguards, is limited to verified organizations.

Gemini 4 Carbon by Google [Nov 10 · agent]
Google's new model for long-running agent work, aimed at complex multi-step software engineering, full-stack web development and security work. Developers can use it on the paid Gemini API tier in Google AI Studio and Android Studio, Google AI Ultra subscribers can try it in AI Studio and Antigravity, and it's rolling out in GitHub Copilot for Pro+, Max, Business and Enterprise plans across VS Code, Copilot CLI, the coding agent, JetBrains and Xcode.

Cursor Projects goes multiplayer [Nov 10 · edit]
Cursor Projects, where one coordinator agent directs a fleet of worker agents on large jobs, now lets you invite teammates. Everyone keeps their own chat with the coordinator, but requirements docs, specs, diagrams and decisions live in shared context every agent can read, a shared task board shows who and which agent is on what along with each task's PR, and teammates' agents can talk to each other to spot overlapping work. Rolling out in beta to all Cursor users; create a Project from the left-hand nav.

Cursor Build Live Apps [Nov 5 · edit]
Paid Cursor plans with cloud agents can now build and publish shareable web apps from a prompt: the agent scaffolds the app, can provision Postgres and Cursor SSO auth, runs security checks before anything goes live, and returns a live link. Team admins see every app and secret; individual plans can share apps publicly.

Boxes.dev [Nov 9 · agent]
A YC launch that moves Claude Code and Codex off your laptop: your local agent ports your dev environment to the cloud, then every agent gets its own machine that can run your full app, test its work end to end and send screenshots. Start agents from the desktop app, your phone, Slack or Linear, forward ports to check work on localhost, and bring your own Claude and Codex subscriptions or API keys, since Boxes doesn't resell tokens.

Air in JetBrains IDEs [Nov 11 · edit]
With the 2026.3 release, JetBrains brings Air into its IDEs as one place to run, track and review coding agents like Junie, Codex, Claude Agent and GitHub Copilot, plus any ACP-compatible agent. Each agent can get its own Git worktree for parallel sessions, presets save your agent, model and instructions, and agents can use the IDE's inspections, run configurations, refactorings, database tools and, in IntelliJ IDEA and CLion, the debugger. Review changes in a panel and leave Code Notes on lines to send back. Works with your existing subscriptions or API keys; a paid JetBrains AI plan isn't required.

Qwen Code 1.0 Preview [Nov 10 · agent]
Alibaba's open-source terminal coding agent folds in the agent capabilities of Qoder, its commercial coding tool. The preview signs in with Qwen or Qoder accounts or any compatible model endpoint, ships native packages for macOS, Linux and Windows that don't need Node.js, and supports resuming sessions and non-interactive runs for scripts and automation. It lives on its own branch while the 0.27 stable line continues, and interfaces may still change.

Runpod for Agents [Nov 10 · deploy]
Runpod expanded its tools for letting coding agents handle GPU infrastructure: point Claude Code, Codex, Cursor or another agent at agents.runpod.io and it installs Runpod skills and connects to a hosted MCP server, with OAuth sign-in where the client supports it. There are native plugins for Claude Code, Codex and ChatGPT, 35 tested skills for picking hardware, deploying Pods and Serverless endpoints, troubleshooting and cost audits, and every tool is marked read-only or destructive so deletions can require your confirmation.

Hermes Agent v0.23.0 [Nov 5 · orchestrate]
NousResearch's "Hearth Release": a new Webapp serves the full Desktop workspace in a browser tab from your own gateway (off by default), and one gateway now owns every session so the CLI, TUI, Desktop, ACP editors, bots and cron attach as clients. Adds /rewind, inline terminal images, pluggable delegation, plugin auto-update, and price tags and spend limits for paid MCP tools; the plugin catalog grew to over 1,050 entries. Breaking: old browser tools replaced by browser_exec and hermes mcp serve renamed to hermes mcp-serve.

Kiro bots in Kiro Crew [Nov 5 · agent]
AWS's Kiro Crew adds generally available bots: named, long-running agents with their own memory, notes, workspace and scoped permissions that start parallel worker sessions, ask for approval when needed, and can run on schedules (for example overnight CI triage or Friday release notes).

All 30 drops, newest first: buildwith.sh/new · RSS

$ diff rankings --since=last-week

+ Context & Indexing joins the board
A new category for tools that help agents understand your codebase: CocoIndex, Serena, Sourcegraph and Repomix.

+ goose enters Coding Agents at #7
The Apache-2.0 agent, now under the Agentic AI Foundation. Kilo Code, Crush, Aider and Trail-code each slip one spot.

+ Kiro enters AI Code Editors at #5
AWS's agentic IDE with spec-driven development joins the editors list.

+ Prices re-checked
Pricing was re-verified on 40+ tools in Wednesday's pass. Composio added its $18/seat Standard team plan, and Neon's free tier now covers up to 100 projects.

60 hand-ranked tools across 11 categories, re-checked every Wednesday. Rankings are never paid. See the board.

$ cat tip.md

Give your agent read-only access first. Runpod's new agent tools mark every call as read-only or destructive, so deletions can ask for your confirmation. Connect your agent to infrastructure like that, let it list and inspect for a day, and only then allow the destructive calls. On Claude Code 2.1.321, `claude remote-control` also takes a --settings file applied to every session it starts, so you can set those guardrails once. More in Agent Ops.

$ cat sponsor.txt

This slot is open. Founding Sponsor weeks start at $99 and include a sitewide bar plus a line in this newsletter. Book a week.

You're getting this because you subscribed at buildwith.sh. Reply to this email to reach us.

$ subscribe --weekly

One email a week: new tools worth your time and ranking changes. No spam.

why subscribe →past issues →

sponsor this spot →