What AI Update Did You Miss Today?

Codex CLI 0.132.0: Python SDK First-Class Authentication & Faster TUI Startup

Codex CLI 0.132.0, released May 20, 2026, delivers first-class authentication support to the Python SDK (openai-codex), allowing developers to programmatically sign in via API key, ChatGPT browser flow, or device-code flow — all without leaving the Python environment. The release also simplifies turn APIs for text-only workflows, adds --output-schema support to codex exec resume for structured JSON output in ongoing automation sessions, and batches terminal capability checks to achieve noticeably faster TUI startup times.

Codex

The Unreasonable Effectiveness of HTML: Why the Claude Code Team Switched from Markdown

Thariq Shihipar, a member of the Claude Code technical staff at Anthropic, published a widely-shared piece arguing that HTML — not Markdown — is the superior output format for Claude Code workflows. The article demonstrates six high-value use cases where HTML's richer expressiveness (interactive elements, SVG diagrams, styled tables, tabbed navigation) delivers dramatically better results than Markdown's flat text limitations. The post sparked major community discussion with over 750,000 views, a Hacker News thread, and coverage in Lenny's Newsletter and Simon Willison's blog.

Claude Code

Gemini 3.5 Flash: Google's Fastest Frontier Model Reaches General Availability

Google launched Gemini 3.5 Flash as generally available on May 19, 2026, at Google I/O 2026, positioning it as the new flagship model for agentic and coding tasks. The model surpasses Gemini 3.1 Pro on nearly all benchmarks — including Terminal-Bench 2.1 (76.2%), MCP Atlas (83.6%), and CharXiv Reasoning (84.2%) — while running 4× faster than competing frontier models. Priced at $1.50/$9.00 per million input/output tokens with a 1M-token context window, Gemini 3.5 Flash is immediately available via the Gemini API, Google AI Studio, and Google Antigravity, and becomes the default model powering the Gemini app and AI Mode in Google Search.

Gemini CLI

Claude Code: Self-Hosted Sandboxes and MCP Tunnels for Managed Agents

At the Code with Claude London event on May 19, 2026, Anthropic announced two infrastructure features for Claude Managed Agents: self-hosted sandboxes (public beta) and MCP tunnels (research preview). Self-hosted sandboxes allow enterprises to run Claude Code's agent tool execution inside their own infrastructure — keeping sensitive files, repositories, and services inside their security perimeter — while the orchestration loop remains on Anthropic's platform. MCP tunnels extend this control to private networks, enabling Claude agents to reach internal databases, proprietary APIs, and ticketing systems through an end-to-end encrypted outbound connection, with no inbound firewall rules or public endpoints required.

Claude Code

Gemini API: Managed Agents and the Antigravity Agent in Public Preview

Google launched Managed Agents in the Gemini API at Google I/O 2026, enabling developers to spin up a fully autonomous, Google-hosted agent with a single API call. The flagship Antigravity Agent (model ID: antigravity-preview-05-2026) — powered by Gemini 3.5 Flash — can plan, write and execute code, browse the web, and manage files inside a persistent Linux sandbox, all via the Interactions API. Custom agents are defined entirely through markdown files (AGENTS.md, SKILL.md), removing the need for custom infrastructure or orchestration code; environment compute is free during the public preview period.

Gemini CLI

Cursor in Jira: Assign Tickets to Cloud Agents

Cursor launched a native Jira integration that lets teams assign Jira work items directly to Cursor's cloud agent, or trigger it via @Cursor mentions in issue comments. The agent reads the ticket's title, description, and comments, then autonomously fixes bugs, adds features, updates tests, or investigates the task — posting progress updates and linking the resulting pull request back in Jira. The integration requires Jira Commercial Cloud with Rovo enabled and is installed through the Cursor integrations dashboard.

Cursor

GitHub Copilot: Gemini 3.5 Flash Now Generally Available

Google's Gemini 3.5 Flash is now generally available in GitHub Copilot, offering near-Pro coding quality at Flash-tier speed and cost. GitHub positions the model as especially well-suited for fast, iterative agentic coding workflows — thanks to strong tool use, fast response times, and efficient caching. The model is accessible to Copilot Pro, Pro+, Business, and Enterprise subscribers across all major IDEs and GitHub Mobile, launching with a 14x premium request multiplier (pricing tentative and subject to change). Enterprise and Business administrators must enable the Gemini 3.5 Flash policy in Copilot settings before it appears for their teams.

GitHub Copilot

GitHub Copilot Code Review: Batch Apply Feedback with Cloud Agent

GitHub has upgraded the workflow for applying Copilot code review suggestions via the cloud agent. The previous "Implement suggestion" button has been redesigned as a full dialog that lets developers choose whether changes go to the current pull request or a new one, select which AI model to use, and add optional guidance. A new "Fix batch with Copilot" button enables addressing multiple review comments in a single cloud agent session, replacing the tedious one-by-one approach for large pull requests.

GitHub Copilot

Lovable: Build Apps on Top of Google's Ecosystem

Lovable expanded its Google connector suite on May 19, 2026, making eight Google service integrations available to all users — timed around Google I/O 2026. The suite covers Gmail, Google Calendar, Google Drive, Google Sheets, Google Slides, Google Maps Platform, Gemini Enterprise, and BigQuery, enabling users to build custom applications directly on top of their existing Google data and workflows. This release adds two connectors not previously available — Google Slides and Google Maps Platform — and broadens access beyond the enterprise tier covered in the April 24 launch. Lovable is also listed as a partner in Google's Agent Gallery inside Gemini Enterprise.

Lovable

Mistral Vibe: Multi-Repository Sessions with --add-dir

Mistral Vibe v2.10.0 introduces the --add-dir flag, enabling developers to pull additional repository roots into a single agent session at startup. This makes it straightforward to work across multiple codebases — a frontend and backend repo, a main application alongside a shared library, or application code with infrastructure configuration — without spinning up separate sessions. The same release also promotes Mistral Vibe's connector integrations to public GA endpoints, marking their first production-grade availability following the beta rollout in April 2026.

Mistral Vibe

claude agents --json: List Live Sessions as JSON for Scripts and Status Bars

Claude Code v2.1.145 adds a --json flag to the claude agents command, enabling developers to programmatically query all live Claude sessions as structured JSON output. This makes it straightforward to build custom integrations — from tmux-resurrect session persistence to status bar widgets and session picker UIs. Additionally, terminal tab titles for claude agents windows now display a count of sessions awaiting input, so developers with multiple parallel agents can see at a glance which ones need attention without switching windows.

Claude Code

Claude Code Plugin Discovery Now Shows Commands, Agents, and Skills Before You Install

Claude Code v2.1.145 upgrades the /plugin Discover and Browse screens to surface a plugin's full manifest — including its commands, agents, skills, hooks, and MCP/LSP servers — before installation. Previously, developers had to install a plugin to discover what it actually contained. This eliminates the install-to-inspect cycle and makes the plugin marketplace meaningfully more browsable.

Claude Code

Mistral Vibe: Live-Editable Plan Mode Display

Mistral Vibe v2.10.0 introduces a live-editable plan display in plan mode, giving developers direct control over the agent's action plan before execution. Previously a read-only view, the plan is now an interactive surface that can be edited directly from the terminal by pressing Ctrl+G. This turns plan mode into a genuine collaborative drafting space, reducing the back-and-forth typically needed to correct a proposed sequence of steps. The same release also adds tool approval debouncing to reduce interruptions during rapid approval flows.

Mistral Vibe

Cursor Composer 2.5: Cursor's Most Capable In-House Coding Model

Cursor released Composer 2.5 on May 18, 2026 — the most capable in-house AI coding model the company has shipped. Built on Moonshot's open-source Kimi K2.5 checkpoint and trained on 25 times more synthetic tasks than its predecessor, Composer 2.5 scores 79.8% on SWE-Bench Multilingual and 63.2% on CursorBench v3.1, matching Anthropic's Claude Opus 4.7 and OpenAI's GPT-5.5 on those benchmarks at roughly one-tenth the cost. The model is designed for long-running, tool-heavy agent sessions and is immediately available inside the Cursor Agent and CLI, with double usage offered to all users for the first week.

Cursor

Lovable MCP Server: Control Lovable from Any AI Client

Lovable launched its official MCP server in research preview, available at https://mcp.lovable.dev for Pro and Business plan users. External AI clients — including Claude Desktop, Claude Code, Cursor, and any MCP-compatible tool — can now connect and manage Lovable projects entirely through natural language. The server exposes project creation, agent messaging, file inspection, SQL execution, analytics queries, and deployment, effectively making Lovable an execution layer within any AI-native workflow. Tool schemas are subject to breaking changes during the preview period.

Lovable