Devin and Cursor broaden enterprise controls as platforms ship multi-agent polish

Enterprise controls land on Devin and Cursor

Two agents took meaningful steps toward making coding agents safer in larger organizations. Devin launched personal automations that run under the user's own identity with a dedicated permission model and badge, and added spend limits per pull request in Devin Review alongside support for GitHub Enterprise Server and GitLab PRs. A V3 API now exposes org-scoped member listing and full automations CRUD. Meanwhile, Cursor Enterprise introduced multi-team organizations — a single umbrella that manages separate security, governance, budget, and model-access controls per team, with unified spend rollup. Users can belong to multiple teams with different roles, and admins move users between teams via dashboard, API, or CSV.

Desktop handoffs and persistence in Codex and Grok Build

OpenAI Codex shipped a notable desktop handoff: the /app command can now transfer a CLI thread into the Codex Desktop on macOS and Windows, so you can start work in the terminal and continue in the GUI without restarting. Codex also brought Sites out in preview — a plugin for building and deploying web apps, dashboards, and games hosted by OpenAI, enabled by default in Business workspaces. The iOS client gained Face ID lock and Windows SSH support.

xAI's Grok Build focused on making long-running workflows survivable: bash commands that hit the timeout are now moved to the background instead of killed, and grok update restarts any older running leader so all clients get the latest binary. A series of patches across several minor releases also fixed false-positive doom-loop terminations on batch failures, grep search timeouts on large repos, and Windows command handling for &-containing commands. Leader sessions gained multi-client live dashboards, and session deletion from the /resume picker.

Fallback models and cross-session hardening in Claude Code

Anthropic's Claude Code introduced a fallbackModel setting that tries up to three fallback models in order when the primary is overloaded — including automatic retry on non-retryable API errors. Cross-session messaging hardened: messages relayed via SendMessage no longer carry user authority, and receivers refuse relayed permission requests. A later release fixed bash failures under Bazel and EDR-protected Go workflows caused by a $TMPDIR override regression, and added version-enforcement managed settings. The 2.1.160 patch fixed a critical bug where restoring a completed session from claude agents dropped chat history and re-ran the original prompt, plus safety prompts before writing to shell startup files and build-tool configs that grant code execution.

Design Mode and agent SDK expansion in Cursor

Cursor doubled down on its canvas and SDK surface. Design Mode added multi-select for adjusting groups of UI components and voice input to narrate changes through the overlay — then arrived in canvases so you can select and annotate rendered UI elements directly. The same update included an interactive context-usage report that breaks down token allocation across system prompts, tools, rules, and skills. The Cursor SDK gained custom tools passed as function definitions, auto-review routing for local tool calls, and nested subagents that can spawn their own subagents — along with JSONL and custom store support for agent metadata persistence.

Smaller waves across the ecosystem

Codex CLI 0.137.0 added enterprise credit limits and multi-agent v2 runtime persistence. Claude Code fixed permission rules and MCP timeouts, and added claude agents --json with waitingFor fields to show what blocks a waiting session. Grok Build shipped image-to-video tools, fixed ghost-cell artifacts in markdown table rendering, and added a plugin install suggestion that appears above the prompt when you type a known marketplace plugin name.