On May 7, Cursor shipped version 3.3. On May 20, it shipped 3.5. Three weeks, two releases, one direction: the IDE is becoming the control plane for an agent-native development organization.
This is a builders-log on both releases combined — because the story they tell together is more important than either one alone.
What Changed in Cursor 3.3 (May 7)
Build in Parallel
The headline feature. When you have a multi-step plan, clicking Build in Parallel causes Cursor to analyze task dependencies and construct a dependency-aware execution graph. Independent steps are dispatched simultaneously to async subagents. Steps that depend on earlier output wait in sequence.
The /multitask slash command triggers the same behavior inline, without opening a plan view.
Technical notes:
- Subagent model is configurable: you can specify a model explicitly, inherit the parent model, or disable subagents entirely
- General aliases are supported (
model: opus,model: sonnet) for subagent config - Real-world performance is codebase-topology-dependent. Loosely coupled projects — separate microservices, independent feature branches — see the biggest gains. Third-party coverage benchmarks 20–30% faster cycles for loosely coupled codebases. Dense monorepos where changes constantly touch shared infrastructure see fewer parallelizable steps and therefore fewer gains
The practical impact: if you have a plan with eight steps and five of them are independent (write tests for module A, write tests for module B, update the readme, add a changelog entry, bump the version), those five now run concurrently. You wait for the two that depend on them, not the full eight.
PR Review in the Agents Window
Cursor has moved the entire PR lifecycle into the Agents Window. Three tabs were added:
- Reviews — inline review threads and top-level PR comments
- Commits — a focused view of commit history for the current PR
- Changes — file tree and changes picker for navigating large diffs
Alongside these: reviewer status banners and pending review banners with quick-action pills. The agent can read review context — a reviewer’s comment, a requested change — and act on it directly inside the editor. No context switching to GitHub or GitLab, no copy-pasting comment text into chat.
The significance is less about any individual tab and more about where this surface lives. The PR review loop now has a home in the same window where the code gets written. The agent sees the review comments in the same context as the codebase.
Split-into-PRs
Available in multitask mode: Cursor analyzes chat context, identifies logical slices in the current changeset, creates a backup snapshot, and proposes a split plan for review before executing.
The snapshot is deliberate. Splitting branches is a destructive operation if something goes wrong; the backup gives you a restore point. When changes have dependencies, Cursor acknowledges them and structures the split accordingly rather than making every slice independent by fiat.
What Changed in Cursor 3.5 (May 20)
Automations in the Agents Window
Cursor Automations previously required going to cursor.com/automations to manage them. They are now also accessible inside the Agents Window in the IDE directly. Both access paths remain; this is about friction reduction.
More important is what happened to the Automations product itself.
Multi-Repo Automations
You can now attach multiple repositories to a single automation. The agent reasons across all attached codebases simultaneously — delivering, testing, and verifying tasks that span multiple repos in a single run.
This is a significant capability shift. Before: automations operated on one repo. A change that needed coordination across a shared library and its consumers required multiple separate automations or manual orchestration. After: one automation, one agent, multiple repos.
No-Repo Automations
A new automation type that has no attached repository. The agent monitors external tools and signals instead of editing code.
Five templates shipped to the Cursor Marketplace:
| Template | What it does |
|---|---|
| Slack digest agent | Summarizes unread DMs and channel messages, prioritized by importance |
| Product analytics agent | Weekly metric digests from data warehouses (Databricks) |
| Product FAQ agent | Monitors Slack channels, auto-responds using docs and context |
| Product finance agent | Extracts financial data from billing providers (Stripe) |
| Customer health agent | Monitors Granola, Slack, Databricks; flags account health shifts |
The no-repo automation is a different category of agent than anything Cursor has shipped before. The prior product was: your code + an AI that helps you write more code. No-repo automations are: your business signals + an AI that monitors them and surfaces what matters.
This is where “IDE as control plane” stops being a metaphor and becomes literal. Cursor is bidding to be the interface through which a development team manages not just code, but the operational signals around it.
Shared Canvases
Teams can now share a link to a live snapshot of a Canvas — the interactive artifacts that agents produce (dashboards, prototypes, data visualizations). Read-only access for recipients; available on Pro, Teams, and Enterprise plans.
The Pattern
Read these two releases together and a strategic intent becomes clear.
3.3 makes the agent faster at executing work (parallel execution) and closes the review loop inside the editor (PR review surface). The agent doesn’t just write code — it participates in the code review cycle.
3.5 extends the agent’s scope beyond a single repository (multi-repo automations) and beyond code entirely (no-repo monitoring templates). The agent doesn’t just participate in your development cycle — it begins monitoring the product and business signals that development responds to.
This is a faster escalation than most people expected. Three weeks between releases. Each release materially extending what “coding assistant” means.
The Cursor Composer 2.5 article we published earlier this month covers the model quality and supply chain story: the underlying model is Kimi K2.5, an open-weight model from Moonshot AI, delivering near-frontier benchmark performance at a fraction of the API cost of Claude Opus 4.7 or GPT-5.5.
The 3.3 and 3.5 releases are the product story on top of that model. Cursor’s bet is not just that the model should be good — it’s that the model should be wired into every part of how a development team works.
What Builders Should Do
Enable parallel execution selectively. Build in Parallel is most effective on plans with genuinely independent steps. If your plans tend to be highly sequential — each step depends on the last — parallel execution adds overhead without much benefit. Try it on feature work before applying it to refactors that touch shared infrastructure.
Use the PR Review surface once per sprint. The value of having reviews, commits, and changes in the Agents Window is context continuity: the agent can see reviewer comments in the same session where it has codebase context. This is most useful for review-driven iteration cycles, not one-pass PRs.
Pilot a no-repo automation for one signal. The Slack digest template is the lowest-stakes entry point. Deploy it, run it for a week, evaluate whether the digest is actually useful. If yes, extend to the analytics or finance templates. The goal is to see whether Cursor as a signal aggregator fits your workflow before committing to it as your operational monitoring surface.
Watch the multi-repo automation carefully. Multi-repo agents are powerful and carry higher blast radius than single-repo agents. Before automating cross-repo work, define the failure mode: what happens if the agent makes a change in the library repo that breaks a consumer repo? Having that answer before the first run is the difference between a useful tool and an incident.
This is a ChatForest builders-log — analysis for AI-native developers making tool decisions. We research and synthesize; we don’t have hands-on access to Cursor 3.3 or 3.5 features directly. All feature descriptions are sourced from the Cursor changelog and community coverage.