MindPattern
← The Wire
Source trail

Cursor — Introducing Composer 2

Public MindPattern findings, entities, and graph evidence that cite this source.

Findings
40
All-time hits
60
High value
33
Last seen
2026-06-29

Connected entities

Cursor BlogCursor 3.6 Ships Auto-Review: Classifier-Gated Autonomous ExecutionTip: Cursor 3's /best-of-n — Run Same Task Across Multiple Models in Parallel Worktrees Then CompareCursor Named Leader in 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents — 70% of Fortune CursorMultiple (Cursor Blog, Google Developers Blog)Cursor 3.5 Ships Multi-Repo Automations and No-Code Agent TemplatesCursor Ships Composer 2.5: Custom Agentic Model Hits 79.8% SWE-bench Multilingual at 10-60x Lower CoPattern: AI Coding Platforms Train Custom Models to Escape Foundation API DependencyCursor 3.4: Cloud Agent Dev Environments with Multi-Repo Support and 70% Faster Cached BuildsMultiplePattern: All Three Major AI Coding IDEs Now Ship Parallel Subagent ExecutionCursor 3.3: Parallel Plan Execution, PR Split, and Subagent Configuration ControlCursor Changelog — SDK ReleaseCursor SDK Public Beta: TypeScript API Turns Cursor's Coding Agent Into Programmable Infrastructure Pattern: IDE Vendors Ship First-Party Security Agents as Standard PR Review Layer

Related findings

  1. 2026-06-29 / SKILLSCursor Composer 2's compaction-in-loop RL cuts mid-task 'forgot the constraint' errors ~50%Composer 2 (released March 19, 2026) trains the model to compress its own ~200K context down to roughly 1,000 tokens mid-session and continue — and because the compression happens inside the RL loop with a reward for maintaining task completion post-compression, the model learns which details to keep, reporting ~50% fewer compaction errors than external summarization. The practitioner signal: 'compaction-in-the-loop' is becoming the differentiator for multi-hour autonomous coding, and external bolt-on summarizers are the weak point. Builders evaluating coding agents for long refactors should test specifically for constraint loss after compression, not just raw context size.
  2. 2026-06-27 / REDDITCursor's June Update: Composer 2.5 Powers Bugbot (~90s Reviews), Pre-Push '/review', and Always-On AutomationsCursor's June release ties its in-house Composer 2.5 model to Bugbot, cutting average review time from ~5 minutes to ~90 seconds while finding ~10% more bugs per review (0.62 vs 0.56) at ~22% lower cost per run. A new /review flow runs Bugbot and Security Review before push, dedups against the eventual PR, and supports incremental 'only what's new' review; Cursor also added Automations — always-on agents plus an /automate skill with GitHub/Slack triggers and computer use. This is a concrete move from chat-IDE to standing review/automation agents in the dev loop.
  3. 2026-06-02 / TOOLSCursor Bugbot Switches to Usage-Based Billing: $1.00-$1.50 per PR Review Starting June 8Cursor is dropping Bugbot's $40/seat/month subscription in favor of usage-based billing for Teams and Individual plans. Average review costs $1.00-$1.50 depending on PR size, with configurable effort levels for deeper or custom review logic. Existing customers transition at their next renewal after June 8. Teams bill from on-demand spend; Individuals from included usage — a significant pricing model shift for AI code review.
  4. 2026-06-01 / TOOLSPattern: Auto-Mode Convergence — Cursor and Claude Code Both Ship Classifier-Based Autonomous ExecutionWithin the same week, both Claude Code (v2.1.158 Auto Mode) and Cursor (3.6 Auto-review) shipped remarkably similar architectures: a classifier subagent reviews each pending action against conversation context and decides allow/retry/ask. Both use allowlists for known-safe operations, sandboxing for containable actions, and model-based judgment for everything else. This convergence suggests classifier-gated autonomy is becoming the industry-standard pattern for balancing agent productivity against safety — neither full manual approval nor unrestricted execution, but a learned middle ground.
  5. 2026-06-01 / TOOLSCursor 3.6 Ships Auto-Review: Classifier-Gated Autonomous ExecutionCursor 3.6 released May 29 introduces Auto-review, a new run mode where allowlisted Shell/MCP/Fetch calls run immediately, sandboxable calls execute in the sandbox, and everything else goes through a classifier subagent that decides allow/retry/ask. Users configure the run mode in Settings > Agents > Run Mode and can steer the classifier with custom instructions. This positions Cursor as having the same classifier-based autonomy pattern Claude Code adopted, creating an industry convergence on gated auto-execution.
  6. 2026-05-29 / SKILLSCursor Ships Jira Integration: @Cursor in Comments Kicks Off Cloud Agents That Fix Bugs, Add Features, and Open PRs from TicketsCursor's Jira integration lets you assign work items to @Cursor or mention it in comments to spawn cloud agents. The agent uses the ticket title, description, and comments plus your repo settings to scope work. When done, it notifies you in Jira and auto-links the PR. Multi-platform steering from Jira, IDE, or Cursor web. Rules can auto-assign certain ticket types. Requires Jira Cloud with Rovo enabled.
  7. 2026-05-29 / TOOLSCursor 3.5: Canvas Sharing, /loop Skill for Long-Running Local Agents, and Jira IntegrationCursor 3.5 (May 20) shipped three notable features beyond the previously-covered 'Build in Parallel': Canvas Sharing lets teams share interactive artifacts (dashboards, reports, interfaces) as live browser links. The /loop skill enables persistent local agents ('check deploy status every 5 minutes', 'work until tests pass'). Jira integration allows agents to pick up tickets, implement fixes, and post PR links back to Jira automatically. Multi-repo automations let agents reason across multiple attached repositories.
  8. 2026-05-28 / TOOLSTip: Cursor's Build in Parallel Identifies Independent Plan Steps and Runs Async SubagentsCursor 3.3's 'Build in Parallel' feature analyzes your plan, identifies independent steps that don't share file dependencies, and spawns async subagents for each — similar to Claude Code's agent teams but integrated into the plan/execute workflow. The key technique: write plans with explicit dependency markers so the parallelizer can split cleanly. Tasks that touch different files or modules parallelize well; tasks sharing a config file or database schema don't. Combine with Cursor's new PR Split to automatically slice parallel work into independent pull requests.
  9. 2026-05-27 / SOURCESCursor Named Leader in 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents — 70% of Fortune 500 Now Using CursorGartner's inaugural Magic Quadrant for Enterprise AI Coding Agents placed Cursor furthest on Completeness of Vision among all vendors. Over 70% of the Fortune 500 now deploy Cursor for coding agents across the software development lifecycle. Cursor is partnering with SpaceXAI to build a frontier model from scratch, alongside in-house training for long-running coding agents via the recently shipped Composer 2.5.
  10. 2026-05-27 / SOURCESCursor Publishes 'What We've Learned Building Cloud Agents' — 50M+ Actions/Day, 40% of Internal PRs from AgentsCursor's Josh Ma details hard-won lessons from scaling cloud agents: development environment fidelity is the single biggest quality factor, they process over 50 million actions per day across 7 million unique workflows, and 40% of Cursor's own PRs now come from cloud agents. They migrated to Temporal for durable execution, going from one to two nines of reliability, and found that decoupling agent loops from machine state enables agents to spawn subagents across pods and outlive their parent processes.
  11. 2026-05-26 / DISPATCHCursor Composer 2.5 Promotional Period Ends — Fast Tier Price Doubles to $3/$15 per Million TokensCursor's promotional double-usage period for Composer 2.5 ended May 25, and the new pricing reveals a significant cost increase on the Fast tier (the default): $3.00/$15.00 per Mtok input/output, up from $1.50/$7.50 in Composer 2. Standard tier pricing held at $0.50/$2.50 per Mtok. Since Fast is the default option most developers use, the effective cost of daily Cursor usage has roughly doubled — a notable development alongside Uber's report of unsustainable AI tooling costs.
  12. 2026-05-25 / TOOLSPattern: AI Coding Model Commoditization — Open-Weight Fine-Tuning Matches Frontier at 1/10th CostCursor's Composer 2.5 proves the commoditization thesis: take an open-weight base (Kimi K2.5, 1T params), fine-tune on proprietary session data, and match Opus 4.7 on SWE-Bench at $0.50/M input tokens vs. $15. Google takes a different route with Gemini 3.5 Flash at 4x speed. The pattern: raw coding capability is no longer a moat — distribution, data flywheel, and agent orchestration are. Builders should expect pricing pressure across all AI coding tools through Q3 2026.
  13. 2026-05-25 / TOOLSCursor Ships Composer 2.5 — In-House Model Built on Kimi K2.5 Matches Opus 4.7 at 1/10th the CostCursor released Composer 2.5 on May 18, a coding model fine-tuned on Moonshot AI's open-weight Kimi K2.5 (1T total params, 32B active MoE) using real-time session data from millions of Cursor users. It hits 79.8% on SWE-Bench Multilingual — matching Opus 4.7 — while pricing at $0.50/M input tokens vs. $15 for frontier models. The Agent Swarm feature decomposes tasks into up to 100 parallel sub-agents with a 1,500 tool-call budget per workflow.
  14. 2026-05-24 / SKILLSCursor 3 Ships /best-of-n: Run the Same Task Across Multiple Models in Isolated Git Worktrees, Compare Outcomes Side-by-Side for Model SelectionCursor 3's /best-of-n command runs identical prompts across multiple models (e.g., Claude Opus, GPT-5.3, Gemini 3.5) each in its own git worktree, then presents outputs side-by-side for selection. Companion /multitask splits requests into async subagents for parallel execution. The May 13 update extended this with cloud-agent environments: multi-repo support, Dockerfile-based config with build secrets, and admin-level governance for running parallelized agent fleets.
  15. 2026-05-24 / TOOLSCursor 3.5 Ships Multi-Repo Automations and No-Code Agent TemplatesCursor 3.5 (May 20, 2026) moves Automations into the Agents Window and introduces multi-repository support — agents can now reason across multiple codebases to deliver, test, and verify tasks in a single automation. A new no-repo automation mode lets agents monitor external tools (Slack, billing, analytics) without any attached codebase. Five new Marketplace templates ship out of the box, including a Slack digest agent and customer health monitor.
  16. 2026-05-23 / AGENTSCursor Ships Composer 2.5: Custom Agentic Model Hits 79.8% SWE-bench Multilingual at 10-60x Lower Cost Than Frontier ModelsCursor released Composer 2.5 on May 18, its own agentic coding model trained with 25x more synthetic tasks than Composer 2 and a novel targeted textual feedback approach. It scores 79.8% on SWE-bench Multilingual (within 1 point of Opus 4.7, ahead of GPT-5.5) and ranks 3rd on the Artificial Analysis Coding Agent Index at $0.50/$2.50 per million input/output tokens — 10-60x cheaper than the two models above it. The update also adds full-screen agent tabs, a floating prompt bar, and configurable tool call density.
  17. 2026-05-21 / SKILLSCursor Composer 2.5 Ships with Substantially Improved Long-Running Task Intelligence and Complex Multi-Step Instruction FollowingReleased May 18, Composer 2.5 is Cursor's model upgrade specifically targeting the failure modes of extended agent sessions: losing context mid-task, misinterpreting compound instructions, and degrading quality on long-running work. Cursor claims 'substantial improvement in intelligence and behavior' with better sustained work quality. Combined with the earlier Cursor 3.0 parallel agent architecture, this makes Cursor's agent mode meaningfully more reliable for the multi-file refactors and feature implementations where previous versions would drift.
  18. 2026-05-19 / TOOLSPattern: AI Coding Platforms Train Custom Models to Escape Foundation API DependencyCursor's Composer 2.5 (built on Kimi K2.5 with custom RL) matches Opus 4.7 at 1/10th token cost. This follows Cursor training Composer 2 and Codex building task-specific models on its own infrastructure. The pattern is clear: every major AI coding platform is investing in purpose-built post-trained models to reduce per-token costs and API dependency. For builders, the 'best model' for coding is increasingly not the general-purpose frontier model — it's a task-specific one.
  19. 2026-05-19 / TOOLSCursor Launches Cloud Agent Dev Environments, Teams Integration, and Bugbot Billing OverhaulCursor's mid-May updates let teams configure full development environments for cloud agents with multi-repo support, enabling parallelized agent fleets handling tasks end-to-end. A May 11 update adds @Cursor mentions in Microsoft Teams channels to delegate tasks to cloud agents. Bugbot shifts from $40/seat/month to usage-based billing effective June 8 renewal, with new customizable effort levels (Default, High, or natural language Custom rules).
  20. 2026-05-18 / SKILLSCursor Ships Bugbot Effort Levels: Configure How Long Your AI PR Reviewer Thinks on Usage-Billed ReviewsCursor's Bugbot (May 11) now supports configurable effort levels for PR reviews under usage billing. Developers can set Bugbot to think longer for deeper reviews on critical PRs or use custom logic to dynamically determine review effort per PR. This is the first AI code review tool to expose a 'thinking budget' dial, letting teams balance review thoroughness against cost on a per-PR basis.
  21. 2026-05-16 / TOOLSCursor Removes Bugbot Seat Fees — Shifts to Usage-Based Billing with Configurable Effort LevelsCursor eliminates per-seat Bugbot fees and transitions to usage-based billing: Teams bill from on-demand spend, Individuals from included usage. Users can now configure review effort (default, high, custom) — high-effort reviews find 0.95 bugs per run on average. Custom logic can dynamically determine effort level per PR based on rules you define.
  22. 2026-05-16 / TOOLSCursor 3.4: Cloud Agent Dev Environments with Multi-Repo Support and 70% Faster Cached BuildsCursor 3.4 (May 13) introduces development environments for cloud agents supporting multiple repositories in a single agent setup, Dockerfile-based configuration with build secrets for secure package registry access, and layer caching delivering 70% faster builds on cache hits. Includes environment version history with rollback, scoped secrets, egress controls, and agent-led environment setup with credential validation.
  23. 2026-05-14 / AGENTSCursor 3.3 Ships Context Usage Breakdown and Persistent Agent Memory Across SessionsCursor 3.3 released May 7 adds two features addressing core agent limitations. Context Usage Breakdown lets developers click an agent's context ring to see exactly how much context is consumed by rules, skills, MCPs, and subagents, enabling diagnosis of context-window issues. Persistent Agent Memory gives agents a notepad (files like MEMORIES.md) that survives between runs where agents decide what to record: progress, patterns, and context. The next session reads those notes and picks up where it left off.
  24. 2026-05-13 / SOURCESCursor Launches Microsoft Teams Integration — @Cursor in Channels to Delegate Coding Tasks and Generate PRsCursor became available in Microsoft Teams on May 11, allowing teams to mention @Cursor in any channel to delegate coding tasks to a cloud agent. Cursor auto-selects the right repository and model based on prompt and recent activity, reads the full thread for context, then implements and creates a PR. New capabilities include always-on Security Reviewer and Vulnerability Scanner agents for PR checks (beta for Teams/Enterprise), and programmatic agents kicked off from CI/CD for summarizing changes and fixing failures. Bugbot switched to usage-based billing.
  25. 2026-05-12 / SKILLSCursor 3.3 Ships Context Usage Breakdown: See Exactly How Much of Your Agent's Context Goes to Rules, Skills, MCPs, and SubagentsCursor 3.3 (May 6) adds a context usage breakdown panel — click the context ring on any agent session to see token allocation across rules, skills, MCP connections, subagents, and conversation history. This is the first IDE to give developers visibility into why their agent is running out of context or behaving erratically mid-session. Practical use: if your rules bucket is consuming 40% of context, you know to trim .cursorrules before blaming the model.
  26. 2026-05-09 / TOOLSPattern: All Three Major AI Coding IDEs Now Ship Parallel Subagent ExecutionAs of May 2026, Claude Code (agent teams with worktrees), Cursor 3.3 (Build in Parallel), and Windsurf (Devin Local with cloud handoff) all support dispatching multiple agent subprocesses simultaneously. The convergence confirms parallel execution as table stakes for AI coding tools — the performance difference between sequential and parallel agent work on multi-file changes is too large to ignore. Subagent model configuration (choose which model runs subagents) is the emerging differentiator.
  27. 2026-05-09 / TOOLSCursor 3.3: Parallel Plan Execution, PR Split, and Subagent Configuration ControlCursor 3.3 (May 7) ships 'Build in Parallel' — click it and Cursor identifies independent parts of your plan and runs them simultaneously using async subagents, keeping dependent steps sequenced. Also adds a built-in PR split workflow that uses chat context to identify logical slices, and lets you configure subagent model selection from settings (e.g., model: opus to always use newest Opus). Includes a new PR review experience with inline threads and commit-by-commit view.
  28. 2026-05-05 / TOOLSCursor Enterprise: Granular Model Controls, Soft Spending Limits, and Team MarketplaceMay 4 release adds admin ability to set allow/blocklists at model and provider level, block entire providers or specific configurations. Soft limits replace hard limits with automatic alerts at 50%/80%/100%. May 1 update lets admins create team marketplace without connecting a repository first, with configurable plugin install behavior.
  29. 2026-05-02 / SKILLSCursor SDK Public Beta: TypeScript API Turns Cursor's Coding Agent Into Programmable Infrastructure with Sandboxed Cloud VMsCursor released a public beta SDK on April 29, 2026 — a TypeScript API letting developers programmatically create, run, and manage Cursor's coding agents from CI/CD pipelines, scripts, or products. Each cloud-configured agent gets a dedicated sandboxed VM with repo clone and dev environment that persists even if the initiating machine goes offline. Rippling, Notion, Faire, and C3 AI are confirmed early adopters wiring agents into automated build-failure triage.
  30. 2026-05-02 / TOOLSPattern: IDE Vendors Ship First-Party Security Agents as Standard PR Review LayerCursor's Security Review beta (May 1) marks a shift: security scanning is moving from third-party CI plugins to first-party IDE features running as always-on agents that review every PR automatically with inline diff comments. Combined with MCP extensibility for existing SAST tools, this signals security-as-default becoming a competitive differentiator for AI coding editors. Expect Windsurf and others to follow within months.
  31. 2026-05-02 / TOOLSTip: Plug SAST and SCA MCP Servers Into Cursor Security Review for Custom Scanning PipelinesCursor's new Security Review accepts custom MCP servers as additional scanning tools. Connect your existing SAST (Semgrep), SCA (Snyk), and secrets scanners (Gitleaks) via MCP, and Cursor's security agents will use them during every PR review alongside its built-in checks. This turns the IDE's security agent into an orchestrator for your existing security toolchain rather than a standalone replacement — the 'bring your own scanner' approach.
  32. 2026-05-02 / TOOLSCursor Team Marketplace Gets Admin Plugin Management With Required/Default PoliciesCursor's May 1 update lets Team and Enterprise admins set up a team marketplace without first linking a repository and manage first-party plugins with three installation behaviors: Default Off (opt-in), Default On (opt-out), and Required (mandatory, non-removable). Plugins bundle MCP servers, skills, subagents, rules, and hooks into single installs — this enables centralized governance over which AI capabilities teams can access.
  33. 2026-05-02 / TOOLSCursor Security Review Beta: Always-On PR Security Agents for Teams and EnterpriseCursor launched Security Review on May 1 with two always-on agent types: Security Reviewer scans every PR for vulnerabilities, auth regressions, prompt injection attacks, and data-handling risks with inline diff comments including severity and remediation, while Vulnerability Scanner runs scheduled codebase scans and posts findings to Slack. Teams can plug in MCP servers for existing SAST, SCA, and secrets scanners to customize the review pipeline.
  34. 2026-04-25 / TOOLSPattern: Agent-First IDE Convergence — Cursor 3, Claude Code, and Codex All Ship Orchestration LayersIn April 2026, all major AI coding environments converged on the same paradigm: managing fleets of agents rather than writing code. Cursor 3.2 ships /multitask with async subagent fleets and multi-root workspaces. Claude Code has agent teams with worktree isolation and background agents. OpenAI Codex runs persistent sandboxed agents. The differentiation is no longer 'can an AI write code' but 'how well can you orchestrate many agents working in parallel across repos.' The editor as a single-agent interface is dead.
  35. 2026-04-25 / TOOLSCursor 3.2 Ships /multitask: Async Subagents, Worktrees, and Multi-Root WorkspacesCursor 3.2 (April 24) introduces /multitask, which spawns async subagents to parallelize queued requests and break large tasks into smaller chunks for simultaneous execution. Multi-root workspaces let a single agent session span frontend, backend, and shared libraries across repos without retargeting. Worktrees improvements let developers run isolated background tasks across branches. This is the most significant agent-orchestration update since Cursor 3.0 launched on April 2.
  36. 2026-04-18 / SKILLSCursor 3 Ships Composer 2: Anysphere's Own Frontier Coding Model at 200+ tok/s, BugBot Hits 80% Resolution Rate, Design Mode Adds Figma-Style UI CanvasCursor 3 (April 2) is Anysphere's biggest release: Composer 2 is a frontier model pretrained from scratch on curated code, then RL-trained against tool-use tasks, priced at $0.50/$2.50 per million tokens. BugBot now learns from PR feedback to create review rules, with resolution rate nearing 80% — 15pp ahead of nearest competitor. Design Mode adds a Figma-style canvas where dragging a component generates JSX and Tailwind. April 13 update added tiled layout and voice input to the Agents Window.
  37. 2026-04-15 / SKILLSCursor 3 Ships Agents Window with Composer 2 on Kimi K2.5 Base: 39% CursorBench Jump, Design Mode, and Cloud-to-Local HandoffCursor 3 (April 2, 2026) replaces the Composer pane with a full-screen Agents Window for parallel agent execution across local, cloud, SSH, and worktree environments. Composer 2, built on Moonshot AI's Kimi K2.5 with 4x-scale RL, scores 61.3 vs 44.2 on CursorBench (39% improvement). Design Mode lets you annotate UI elements in-browser for precise visual feedback instead of describing changes in text. Worktree-based execution runs the same prompt across multiple models for side-by-side comparison. Cloud-to-local handoff lets you start a task in the cloud and pull results back seamlessly.
  38. 2026-04-12 / TOOLSTip: Cursor 3 Design Mode — Annotate UI Elements in Browser for Precise Agent FeedbackCursor 3's Design Mode in the Agents Window lets you annotate and target specific UI elements directly in the browser, then feed those visual annotations to the agent as context. Instead of describing UI issues in natural language, you click and annotate elements visually, reducing ambiguity in frontend iteration. This is especially effective for CSS and layout corrections where spatial reference matters more than textual description.
  39. 2026-04-12 / TOOLSTip: Cursor 3's /best-of-n — Run Same Task Across Multiple Models in Parallel Worktrees Then CompareCursor 3's /best-of-n command runs the same task in parallel across multiple models, each in its own isolated git worktree, then displays outcomes side by side for comparison. This enables direct head-to-head evaluation of different models on your actual codebase rather than relying on synthetic benchmarks. Combine with Agent Tabs to view all parallel runs simultaneously in a grid layout, and /worktree for isolated single-model experiments.
  40. 2026-04-11 / TOOLSCursor Bugbot Now Self-Improves from PR Feedback in Real Time — Highest Resolution Rate to DateCursor shipped a Bugbot update that enables the review bot to learn from developer feedback on pull requests and convert those signals into learned rules that improve future reviews. The update also adds MCP support for Bugbot and improvements to Bugbot Autofix. Cursor claims this is its highest resolution rate to date — meaningful for teams using Cursor as their primary agent-driven code review layer.
Open latest cited source