TL;DR: Cursor 3's Agents Window now supports tiled layouts with up to 50 parallel agents per team, making multi-repo orchestration practical for the first time. Claude Code shipped three rapid-fire versions (2.1.101 through 2.1.107) tightening enterprise terminal workflows. Bugbot's resolution rate has climbed past its 52% beta baseline. Here's what shipped, ranked by impact, and what you should do about it.

Cursor 3 launched April 2, 2026, and the headline feature isn't a smarter autocomplete. It's a fundamentally different mental model for how engineering work gets done.

The new Agents Window now supports a tiled layout where you can run up to 10 parallel agents per user, or 50 per team. This isn't a cosmetic change. It's the difference between having one contractor on a job site and running a coordinated crew. In practice, a single engineer can now spin up parallel agents tackling separate repos simultaneously, monitor their progress in a unified tiled view, and intervene where needed. Sessions have been clocked at 170 files edited per run. That's not a coding assistant. That's a force multiplier. The architectural implication: your best engineers stop being individual contributors and start being agent orchestrators. Their job is to set direction, review output, and handle the judgment calls that agents still get wrong. If you haven't restructured how your senior engineers spend their time to account for this, you're leaving productivity on the table.

Cursor 3 also shipped multi-repo support inside the Agents Window and cloud agents that run on self-hosted infrastructure. The self-hosted angle matters enormously for enterprise teams who've been blocked on Cursor adoption by data residency or security review requirements. If that's been your blocker, it's gone now. Multi-repo support means agents can reason across service boundaries, which is where most real-world complexity lives. Microservice architectures, shared libraries, cross-repo refactors: Cursor's agents can now work at the systems level, not just the file level.

Cursor Composer 2, which launched March 19, 2026, runs on a Kimi K2.5 base with 4x-scale reinforcement learning. The RL training checkpoints every 5 hours from live user interactions, meaning the model is continuously adapting to real workflows, not just synthetic benchmarks. Performance improvements show up on Terminal-Bench 2.0 and SWE-bench Multilingual, but the more interesting detail is that the model is self-improving from your team's actual usage patterns. Bugbot has also quietly become more useful. Launched at 52% resolution rate in its July 2025 beta, it has since climbed via learned rules extracted from PR feedback. The mechanism matters here: Bugbot isn't just applying static heuristics. It's building rules from how your team responds to its suggestions, which means it gets more accurate the longer you use it. Teams that adopted it early now have a compounding advantage over teams that didn't. One more Cursor note worth flagging: the company hit $2B ARR in January 2026 following a Series D at a $29.3B valuation backed by Google and NVIDIA. At that scale, the product roadmap has real resources behind it. The pace of shipping reflects that.

Anthropic shipped versions 2.1.101, 2.1.105, and 2.1.107 of Claude Code in rapid succession. None of these are headline grabbers individually, but together they signal a clear strategic focus: terminal-first teams, enterprise environments, and extensibility for power users.

Version 2.1.105 added a `path` parameter to `EnterWorktree` and introduced the PreCompact hook. The `PreCompact` hook is the more significant of the two: it allows you to block context compaction before it happens, giving teams the ability to inject custom logic into how Claude Code manages long-running sessions. For teams building automation pipelines on top of Claude Code, this is the kind of low-level control that makes the difference between a tool you use and a tool you build on.

The `EnterWorktree` path parameter is a workflow quality-of-life improvement for teams using Git worktrees heavily, letting agents switch working contexts more precisely.

Version 2.1.101 added a `/team-onboarding` command and OS CA certificate trust. The CA certificate trust is the enterprise unlock: if your network routes traffic through a corporate proxy, Claude Code now respects your system's certificate store, removing a friction point that was blocking deployment in enterprise environments. The `/team-onboarding` command is exactly what it sounds like: a structured way to get new engineers up to speed with Claude Code in a team context. Early-stage adoption friction is a real tax on productivity, and building onboarding into the tool itself is the right call.

Version 2.1.107 surfaced thinking hints earlier during long operations. This is a UX improvement, not a capability change. That said, visibility into what an agent is doing mid-operation reduces the "is this frozen?" anxiety that causes engineers to interrupt long-running tasks prematurely. Less interruption means better outputs from long operations.

Two different bets on how AI-native engineering works. Cursor is building an agent orchestration platform with an IDE interface. Claude Code is building a terminal-native agent environment with enterprise extensibility hooks. Neither approach is wrong. They suit different team archetypes.

The practical takeaway: if your team lives in the terminal and operates in regulated enterprise environments, Claude Code's 2.1.x series is directly addressing your blockers. If your team is building on a greenfield or modern stack and wants to move toward agent orchestration at scale, Cursor 3 is the more aggressive bet. The most capable teams in 2026 are running both. Cursor for agentic output generation and parallel task execution, Claude Code for terminal automation pipelines and enterprise integration. Treating them as mutually exclusive is a false choice.

The IDE is becoming an anachronism as a frame for what these tools are. Cursor 3's Agents Window isn't a better editor. It's a coordination layer for a small team of autonomous agents that your engineers manage. Claude Code's hook architecture isn't better autocomplete. It's infrastructure for building custom AI workflows that fit your specific engineering processes. The teams winning right now aren't the ones with the most engineers. They're the ones who figured out soonest how to operate AI agents at the systems level rather than the file level. Cursor at 50 parallel agents per team and Claude Code's `PreCompact` hooks are both pointing in the same direction: the unit of engineering work is shifting from individual commits to orchestrated, multi-agent campaigns across entire codebases. The engineers who understand that shift, who can architect agent workflows, set meaningful constraints, and review outputs at scale, are the ones who will be worth 10x what a file-by-file coder is worth. Finding those engineers is harder than it's ever been. That's exactly the problem worth solving.