<p><strong>TLDR:</strong> Anthropic shipped multi-agent code review for Claude Code today, dispatching parallel agents across your full codebase on every PR open. The feature is in research preview for Team and Enterprise users, costs $15-25 per PR, and Anthropic’s internal numbers show substantive review rates jumping from 16% to 54%. Meanwhile, VS Code 1.110 quietly landed a set of features that together look less like IDE updates and more like an agent orchestration layer. JetBrains also launched Air into public preview, a new environment built around running multiple agents concurrently rather than around the editor itself.</p>
<hr>
<h2>Claude Code multi-agent code review</h2>
<p>When a PR opens on GitHub, the system spawns a fleet of parallel agents that traverse the entire codebase, not just the diff, to find logic bugs. Each finding gets verified before surfacing, and issues are ranked by severity. Anthropic’s internal numbers: 16% to 54% substantive review rate, fewer than 1% of flagged issues marked incorrect by engineers, and on PRs with 1,000+ changed lines, 84% get actionable findings averaging 7.5 issues. Real-world catches include a type mismatch in TrueNAS’s ZFS encryption refactoring and an auth-breaking one-liner in a production service. The legitimate concern worth sitting with: when the same AI product writes and reviews code, you get a confirmation loop. Anthropic’s multi-agent verification step is meant to address that, but teams will have to judge for themselves whether it actually does.</p>
<h2>VS Code 1.110 ships agent governance primitives</h2>
<p>The headline features in 1.110 are agent plugins and lifecycle hooks. Plugins bundle slash commands, MCP servers, custom agents, and hooks into Git-distributable packages, so platform teams can version-control and enforce AI agent behavior across the org. Hooks let you attach shell commands to agent lifecycle events — session start, stop, tool calls — which means deterministic policy enforcement that doesn’t depend on the model making the right call. Other additions: mid-run guidance so you can redirect a running agent without waiting for it to finish, conversation forking via <code>/fork</code>, manual context compaction via <code>/compact</code>, and a built-in Explore subagent for parallelized codebase research. Taken together this is less a feature release and more a governance layer arriving before most teams have thought to ask for one.</p>
<h2>JetBrains Air goes into public preview</h2>
<p>Air is JetBrains’ new environment designed around running multiple agents concurrently — Codex, Claude Agent, Gemini, and their own Junie — within a single interface. The core idea is giving agents proper context: you reference specific lines, commits, or methods when defining a task rather than pasting snippets into a chat window. Agents run locally or in isolated Docker containers and Git worktrees to keep experimental work sandboxed. Currently macOS-only, free to download, works with a JetBrains AI subscription or your own API keys.</p>
<h2>Fake Claude Code install pages are spreading infostealer malware</h2>
<p>Push Security found a campaign called “InstallFix” cloning Anthropic’s Claude Code installation pages and serving them through Google-sponsored search ads. The fake pages swap the legitimate install command for a script that deploys the Amatera infostealer. On macOS it runs a base64-obfuscated second-stage binary; on Windows it uses <code>mshta.exe</code> to stay under the radar. The attack works because copy-pasting a one-liner install command from a website is completely normal developer behavior. Don’t run terminal commands copied from sites you landed on through a search ad, and verify the domain matches official docs before copying anything.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>GitHub Copilot + Jira integration:</strong> Copilot’s coding agent can now be assigned Jira issues directly and will open a draft PR in GitHub, posting progress updates back to Jira without you switching tools. Requires Jira Cloud with Rovo and a paid Copilot plan.</li>
<li><strong>Datadog MCP Server GA:</strong> Now generally available and connecting Datadog observability to Claude Code, Cursor, Codex, and Copilot; documented use cases include automated decommissioning of unused services and incident response agents correlating alerts with recent feature flag changes.</li>
<li><strong>Crafting raises $5.5M seed:</strong> The startup gives AI coding agents access to production-like environments with real dependencies and infrastructure rather than sandboxes; early customers include Brex, Faire, and Webflow, with teams reporting 25% more PRs shipped quarter over quarter.</li>
<li><strong>Intercom raises $250M for Fin AI at $2B+ valuation:</strong> Notably, they ditched GPT-4 for fine-tuned Qwen models on some tasks to save $250K/month.</li>
<li><strong><a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> files may hurt more than help:</strong> New research recommends omitting LLM-generated context files entirely and limiting human-written instructions to details the model can’t infer on its own.</li>
<li><strong>Perplexity Computer now runs Claude Code as a subagent:</strong> Demonstrated picking up an open GitHub issue, writing the fix via Claude Code, and submitting a PR via GitHub CLI without human input.</li>
<li><strong>Terminal Use (YC W26) launches:</strong> Positions itself as “Vercel for filesystem-based agents” with built-in Claude Agent SDK and Codex SDK support, persistent workspaces across turns, and CI/CD pipelines.</li>
<li><strong>GitHub Copilot hits 60M code reviews:</strong> Announced alongside GPT-4.5 GA in Copilot and a new agentic review architecture with built-in security scanning.</li>
<li><strong>Codex OSS program:</strong> Open-source maintainers can apply for API credits and six months of ChatGPT Pro with Codex; the weekly rate limit also reset today.</li>
<li><strong>GitHub Copilot cancellations at startups:</strong> Gergely Orosz notes companies aren’t restricting AI tool use — devs already migrated to Claude Code, Codex, and Cursor agents, so Copilot licenses just go unused.</li>
<li><strong>Prompt caching in Claude Code:</strong> A 30-minute coding session achieves 92% cache hit rate and 81% cost reduction by separating static context from dynamic conversation history; the key rule is never modify tool definitions mid-session.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Anthropic launched multi-agent code review for Claude Code, dispatching parallel agents across entire codebases on PR open. Internal data shows substantive review rates jumping from 16% to 54%, with fewer than 1% of flagged issues marked incorrect. The feature costs $15-25 per PR and is in research preview for Team and Enterprise users. VS Code 1.110 shipped agent governance primitives including plugins, lifecycle hooks, mid-run guidance, and conversation forking — effectively turning the IDE into an agent orchestration layer. JetBrains launched Air into public preview, a multi-agent environment supporting Codex, Claude, Gemini, and Junie running in sandboxed Docker containers. A security warning: fake Claude Code install pages are spreading the Amatera infostealer via Google-sponsored ads. Additional notable items include GitHub Copilot's Jira integration, Datadog MCP Server GA, prompt caching achieving 81% cost reduction in Claude Code, and research suggesting AGENTS.md files may hurt more than help.

Claude Code review launches, VS Code becomes an agent control plane