getagentictools

2026-07-08

This Week in Agentic Coding #1

TLDR

  • Claude Sonnet 5 launched: performance close to Opus 4.8, $2/$10 per Mtok until August 31, now the default in Claude Code.
  • Anthropic redeployed Fable 5 globally and proposed an industry-wide jailbreak severity framework.
  • OpenAI published a preview of GPT-5.6 Sol, its next-generation model.
  • Cursor shipped an iOS app for cloud agents and team-managed MCP marketplaces.
  • Cline 4.0 had a rocky week: a rollback release, then Sonnet 5 support.

Models & APIs

Claude Sonnet 5 is the new default workhorse (announcement)

Anthropic calls Sonnet 5 "the most agentic Sonnet model yet", with performance "close to that of Opus 4.8, but at lower prices". Introductory pricing is $2 per million input tokens and $10 per million output through August 31, then $3/$15. It is the default model in Claude Code as of v2.1.197, which lists a native 1M-token context window. June 30.

Fable 5 is back online worldwide (announcement)

Anthropic is returning Fable 5 to global availability as of July 1, after the June suspension. Alongside the redeployment, the company proposed an industry-wide framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other partners. June 30.

OpenAI previews GPT-5.6 Sol (announcement)

OpenAI published a first look at GPT-5.6 Sol, which it positions as a next-generation model. Details are thin so far; the preview post is the primary source, and the HN thread carries the speculation. June 26.

Agentic Coding Tools

Claude Code shipped six releases this week (changelog)

Versions 2.1.195 through 2.1.200 landed. The highlights: Sonnet 5 became the default model (v2.1.197), Claude in Chrome went GA and subagents now run in the background by default (v2.1.198), stacked slash-skill invocations like /skill-a /skill-b now load up to 5 skills (v2.1.199), and the "default" permission mode was renamed to "Manual" across the CLI and IDE plugins (v2.1.200).

Cursor puts cloud agents in your pocket (changelog)

Cursor for iOS is in public beta on all paid plans. You launch and manage always-on cloud agents from the phone the same way you would on desktop, with any frontier model, and you can describe ideas by voice. June 29.

Cline 4.0 ships, stumbles, recovers (releases)

Cline rolled its stable VS Code extension back to the pre-SDK-migration codebase in v4.0.1 to resolve regressions reported in 4.0.0, shipping the old code under a higher version number while the migration continues on main. By v4.0.5 the week improved: Claude Sonnet 5 support landed across the Anthropic, Bedrock, Vertex, Claude Code, OpenRouter, and Vercel AI Gateway providers. June 28 to July 2.

Kimi K2.7 Code rolls into GitHub Copilot (changelog)

The open-weight model is rolling out to Copilot Pro, Pro+, and Max plans, with Business and Enterprise following in the coming weeks, off by default until an admin enables it. It is billed at provider list pricing under usage-based billing, making it a lower-cost option in the picker. July 1.

MCP Ecosystem

Cursor makes MCP a team-managed resource (changelog)

Admins can now configure Team MCP servers once and distribute them across cloud agents, the agents window, IDE, and CLI through team marketplaces. Members install approved integrations locally without touching server config, and marketplace access can be restricted by organization group. June 30.

Claude Code closes an MCP self-approval hole (changelog)

As of v2.1.196, claude mcp list and claude mcp get no longer spawn .mcp.json servers that a repo self-approved through a committed settings file. Worth knowing if you audit unfamiliar repos with MCP servers checked in.

📊 New in the Directory

The digest launches this week, so instead of movers, here is what currently tops each category, by real usage:

brainstorming (skill): structured design dialogue that validates ideas before implementation. 215,900 stars on its home repo.

Context7 (MCP server): pulls current library documentation into context instead of training-data memories. 57,464 stars.

Respond To Discussion Features (loop): Payload CMS runs this to find GitHub discussions resolved by recent releases and draft responses. 43,332 stars.

everything-claude-code (plugin): skills, instincts, and configs for tuning the whole agent setup. 141,890 stars.

Browse all at getagentictools.com.

âš¡ Worth Your Time

Simon Willison shows how to have your agent record video demos of its own work with shot-scraper video. Point your closing loop at it and get a screen recording as the proof artifact. June 30.

Be pragmatic, not dogmatic.

Someone forwarded you this? Get The Agentic Coding Playbook, free.

Want to reach thousands of developers building with AI? Sponsor the digest.