Agent Pipeline
Install from VS Marketplace | Install from Open VSX (Cursor)
Multi-agent pipeline orchestrator for Cursor IDE. Automates the software development workflow through 4 specialized AI agents: Planner, Implementer, Reviewer, and Tester.
Each agent has a specific role, receives auto-generated context-rich prompts, and passes its output to the next agent in the chain — creating a structured, reviewable development process.
Table of Contents
Pipeline Flow
Start → Planner → [Review] → Implementer → [Review] → Reviewer → [Review] → Tester → Done
Each agent phase follows this cycle:
- Extension generates a prompt with task description + project context + previous agent outputs
- Prompt is automatically copied to your clipboard
- A notification tells you which AI model to select in Cursor Chat
- You paste the prompt into Cursor Chat (
Cmd+L)
- The AI works according to the agent's
.mdc rules and prompt instructions
- The AI saves its output via the
pipeline_save_output MCP tool (or you can complete manually)
- You review the output in the sidebar and approve or reject
- On approve: pipeline advances to the next agent. On reject: current agent retries with your feedback
Agents
Planner
Role: Analyzes the task and creates a detailed implementation plan.
What it produces:
- Overview of what needs to be done
- List of files to create, modify, or delete
- Ordered implementation steps with dependencies
- Data flow diagrams
- Risk analysis and mitigations
- Acceptance criteria
Default model: claude-4.5-sonnet
Implementer
Role: Writes the actual code following the Planner's output.
What it produces:
- New files and code changes as specified in the plan
- Summary of all files created and modified
- List of added dependencies
- Notes on any deviations from the plan
Default model: claude-4.6-opus
Tip: Claude 4.6 Opus is the most capable model and is recommended for complex implementations. For simpler tasks, claude-4.5-sonnet offers a good balance of speed and quality.
Reviewer
Role: Performs a code review of the Implementer's work. Does NOT write or modify any code.
What it produces:
- Overall assessment: APPROVE or REQUEST_CHANGES
- Summary of findings
- Issues found with severity (Critical / Major / Minor), file location, and fix suggestion
- Positive observations
- Test recommendations for the Tester agent
Review criteria:
- Correctness — Does the code match the plan?
- Code Quality — Clean code, DRY, proper naming
- Security — No vulnerabilities, proper input validation
- Performance — No obvious bottlenecks
- Error Handling — Edge cases covered
- Readability — Well-documented and understandable
Default model: gpt-5.2
Note: The Reviewer prompt explicitly instructs the AI to only read and evaluate — never create or modify files. If the Reviewer tries to implement code, reject and retry.
Tester
Role: Creates and runs tests for the implementation.
What it produces:
- Test files (unit tests, integration tests)
- Test summary — number of tests and what they cover
- Coverage notes — areas covered and gaps
- Run instructions — how to execute the tests
Testing strategy:
- Unit tests for individual functions/methods
- Integration tests for component interactions
- Edge case and boundary condition tests
- Regression tests to ensure nothing is broken
Default model: claude-4.5-sonnet
Model Configuration
Each agent can use a different AI model. This lets you optimize cost and quality — for example, use a fast model for planning and a powerful model for implementation.
Changing Models via Settings
- Open Settings:
Cmd + ,
- Search for
agentPipeline.models
- Set the model for each agent:
| Setting |
Description |
Default |
agentPipeline.models.planner |
Model for the Planner agent |
claude-4.5-sonnet |
agentPipeline.models.implementer |
Model for the Implementer agent |
claude-4.6-opus |
agentPipeline.models.reviewer |
Model for the Reviewer agent |
gpt-5.2 |
agentPipeline.models.test |
Model for the Tester agent |
claude-4.5-sonnet |
Available Models
Current Models (Feb 2026)
| Model |
Provider |
Context |
Notes |
claude-4.6-opus |
Anthropic |
200k / 1M max |
Most capable — recommended for implementation |
claude-4.5-sonnet |
Anthropic |
200k / 1M max |
Fast and capable — great all-rounder |
composer-1.5 |
Cursor |
200k |
Cursor's built-in model |
gemini-3-pro |
Google |
200k / 1M max |
Google's most capable model |
gemini-3-flash |
Google |
200k / 1M max |
Fast Google model |
gpt-5.3-codex |
OpenAI |
272k |
OpenAI's latest coding model |
gpt-5.2 |
OpenAI |
272k |
OpenAI flagship |
grok-code |
xAI |
256k |
xAI coding model |
Previous Generation (BYOK)
| Model |
Provider |
Notes |
claude-3.5-sonnet |
Anthropic |
Previous gen fast model |
claude-3-opus |
Anthropic |
Previous gen most capable |
claude-3-haiku |
Anthropic |
Previous gen fastest |
gpt-4o |
OpenAI |
Previous gen flagship |
gpt-4o-mini |
OpenAI |
Previous gen lightweight |
gpt-4-turbo |
OpenAI |
Previous gen turbo |
gpt-4 |
OpenAI |
Previous gen base |
gemini-1.5-pro |
Google |
Previous gen pro |
gemini-1.5-flash |
Google |
Previous gen flash |
deepseek-v3 |
DeepSeek |
Open-source model |
deepseek-r1 |
DeepSeek |
Open-source reasoning |
o1-mini |
OpenAI |
Reasoning model (mini) |
o1-preview |
OpenAI |
Reasoning model |
Note: You can also type any custom model name that Cursor supports. The dropdown is a convenience — not a limitation.
Where Models Appear
- Sidebar: Each agent step shows its configured model as a badge
- Prompt: The generated prompt includes a model reminder at the top
- Notification: When copying a prompt, the notification tells you which model to select
- Settings button: The sidebar has a "Configure Models" button that opens Settings directly
Recommended Configurations
Cost-optimized:
| Agent | Model |
|-------|-------|
| Planner | gemini-3-flash |
| Implementer | claude-4.5-sonnet |
| Reviewer | composer-1.5 |
| Tester | gemini-3-flash |
Quality-optimized (default):
| Agent | Model |
|-------|-------|
| Planner | claude-4.5-sonnet |
| Implementer | claude-4.6-opus |
| Reviewer | gpt-5.2 |
| Tester | claude-4.5-sonnet |
Maximum power:
| Agent | Model |
|-------|-------|
| Planner | claude-4.6-opus |
| Implementer | claude-4.6-opus |
| Reviewer | claude-4.6-opus |
| Tester | gpt-5.3-codex |
Installation
From VSIX (Recommended)
- Get the
.vsix file (e.g., agent-pipeline-0.4.0.vsix)
- In Cursor:
Cmd+Shift+P → "Extensions: Install from VSIX..."
- Select the
.vsix file
- Reload Cursor (
Cmd+Shift+P → "Developer: Reload Window")
On first activation, the extension will automatically:
- Configure the MCP server in
.cursor/mcp.json
- Prompt you to install agent rule files (
.mdc) in .cursor/rules/
From Source (Development)
cd extension
npm install
npm run compile
# Press F5 to launch Extension Development Host
Getting Started
Step-by-step Walkthrough
- Open the sidebar: Click the pipeline icon in the Activity Bar (left side)
- Start a pipeline: Click "Start Pipeline" and describe your task
- Example: "Add user authentication with JWT tokens and refresh token support"
- Planner phase:
- A prompt is auto-copied to your clipboard
- Select the model shown in the notification (e.g.,
claude-3.5-sonnet)
- Open Cursor Chat (
Cmd+L) and paste the prompt
- The AI will analyze the task and produce a plan
- When done, the sidebar updates to "Review Plan"
- Review the plan:
- Read the Planner's output in the sidebar (expand "Agent Outputs")
- Click "Approve & Continue" if satisfied, or "Reject & Retry" with feedback
- Implementer phase:
- Same flow: copy prompt → select model → paste in Chat → AI implements the code
- The AI follows the Planner's output to write code
- Reviewer phase:
- The Reviewer reads the plan and implementation, then produces a review report
- It does NOT modify any files — only reviews
- Tester phase:
- The Tester creates tests based on the plan, implementation, and review
- Done! All phases completed.
Auto-Complete on File Save
By default, the extension watches for file saves while an agent is active. When files stop being saved (after a 5-second debounce), the phase is automatically completed and you get a notification to review.
This means you don't need to manually click "Complete Phase" — the extension detects when the AI is done by observing file save activity.
You can configure this in Settings:
agentPipeline.autoComplete.enabled — Enable/disable auto-complete (default: true)agentPipeline.autoComplete.debounceSeconds — Seconds to wait after last save (default: 5, range: 2-30)
Sometimes the AI may not automatically call pipeline_save_output. If a phase gets stuck on "Working..." and auto-complete doesn't trigger:
- Click "Complete Phase" in the sidebar
- Optionally paste the AI's output into the input box
- The pipeline advances to the review step
Commands
All commands are available via Cmd+Shift+P:
| Command |
Description |
Agent Pipeline: Start Pipeline |
Begin a new pipeline with a task description |
Agent Pipeline: Approve & Continue |
Approve current phase output and advance to the next agent |
Agent Pipeline: Reject & Retry |
Reject output and retry the current phase (with optional feedback) |
Agent Pipeline: Copy Current Prompt |
Re-copy the active agent's prompt to clipboard |
Agent Pipeline: Complete Current Phase |
Manually complete a phase if the AI didn't call the MCP tool |
Agent Pipeline: Reset Pipeline |
Clear all pipeline state, outputs, and start fresh |
Agent Pipeline: Setup Rule Files |
Install or update .mdc agent rules in the workspace |
The sidebar shows:
- Pipeline Steps — Visual progress indicator with status for each agent (Pending / Working / Awaiting Review / Done) and the configured model badge
- Task — The current task description
- Current Phase — Active phase with status badge and instructions
- Agent Outputs — Expandable section showing each agent's output (click to expand/collapse)
- Action Buttons:
- Start Pipeline — visible when idle
- Copy Prompt — visible when an agent is active
- Complete Phase — visible when an agent is active (manual completion)
- Approve & Continue — visible during review phases
- Reject & Retry — visible during review phases
- Reset — visible when pipeline is running
- Configure Models — always visible, opens model settings
Settings
All settings are under the agentPipeline namespace:
| Setting |
Type |
Default |
Description |
agentPipeline.autoComplete.enabled |
boolean |
true |
Auto-complete phase when files are saved |
agentPipeline.autoComplete.debounceSeconds |
number |
5 |
Seconds to wait after last file save before auto-completing |
agentPipeline.models.planner |
string |
claude-4.5-sonnet |
AI model for the Planner agent |
agentPipeline.models.implementer |
string |
claude-4.6-opus |
AI model for the Implementer agent |
agentPipeline.models.reviewer |
string |
gpt-5.2 |
AI model for the Reviewer agent |
agentPipeline.models.test |
string |
claude-4.5-sonnet |
AI model for the Tester agent |
Settings can be configured at:
- User level — applies to all workspaces
- Workspace level — applies only to the current project (overrides user settings)
The extension includes an MCP server that exposes tools for the AI to interact with the pipeline. These are called automatically by the AI when working within a pipeline phase.
| Tool |
Description |
pipeline_get_task |
Returns the current task description and pipeline status |
pipeline_get_outputs |
Returns outputs from previous agents (by agent name) |
pipeline_save_output |
Saves the current agent's output and triggers phase transition |
pipeline_get_context |
Returns project file tree and detected tech stack |
pipeline_get_status |
Returns full pipeline status including phase history |
The MCP server is automatically configured in .cursor/mcp.json when the extension activates. It uses an absolute path to the server bundled inside the extension, so it works regardless of which workspace you open.
Rule Files (.mdc)
The extension ships with 5 rule files that tell Cursor's AI how to behave as each agent:
| File |
Purpose |
global.mdc |
Shared rules — always active, defines the pipeline concept and MCP tool usage |
planner.mdc |
Planner agent behavior — task analysis, plan structure, output format |
implementer.mdc |
Implementer agent behavior — coding guidelines, output format |
reviewer.mdc |
Reviewer agent behavior — review-only (no code changes), checklist, output format |
test.mdc |
Tester agent behavior — testing strategy, frameworks, output format |
These are installed to .cursor/rules/ in your workspace. You can customize them to fit your project's needs.
Reinstalling Rules
If you modify the rules and want to restore the originals:
Cmd+Shift+P → "Agent Pipeline: Setup Rule Files"
Building from Source
Prerequisites
Build
# Install dependencies
npm install
# Build extension + MCP server
npm run build:all
# Package as VSIX
npx vsce package
This creates agent-pipeline-0.5.0.vsix ready for distribution.
Development
# Build in development mode
npm run compile
# Watch for changes
npm run watch
# Press F5 in Cursor/VS Code to launch Extension Development Host
Project Structure
extension/
├── src/
│ ├── extension.ts # Main entry point, commands, auto-setup
│ ├── pipeline/
│ │ ├── AgentConfig.ts # Agent definitions, phases, model defaults
│ │ ├── PipelineManager.ts # State machine, phase transitions
│ │ └── PromptGenerator.ts # Context-rich prompt generation per agent
│ ├── utils/
│ │ ├── storage.ts # Pipeline state persistence (.agent_pipeline/)
│ │ └── fileAnalyzer.ts # Workspace analysis (file tree, tech stack)
│ ├── views/
│ │ └── SidebarProvider.ts # Webview sidebar (inline HTML/CSS/JS)
│ └── mcp/
│ └── server.ts # MCP server exposing pipeline tools
├── resources/
│ ├── icons/pipeline.svg # Activity bar icon
│ └── rules/*.mdc # Agent rule files
├── package.json # Extension manifest, settings, commands
├── webpack.config.js # Extension bundle config
├── webpack.mcp.config.js # MCP server bundle config
└── tsconfig.json # TypeScript config
Troubleshooting
- Make sure you reloaded Cursor after installing the VSIX (
Cmd+Shift+P → "Developer: Reload Window")
- Check that the pipeline icon appears in the Activity Bar (left side)
Agent phase stuck on "Working..."
- The AI may not have called the
pipeline_save_output MCP tool
- Click "Complete Phase" in the sidebar to manually advance
- You can paste the AI's output into the input box, or leave it empty
Reviewer is writing code instead of reviewing
- This was fixed in v0.4.0 with stronger prompt instructions
- If it still happens, click "Reject & Retry" — the retry prompt reinforces the review-only role
- Check that
reviewer.mdc in .cursor/rules/ contains the "DO NOT write code" instructions
MCP server not connecting
- Check
.cursor/mcp.json in your workspace — it should contain an agent-pipeline entry
- Run
Cmd+Shift+P → "Agent Pipeline: Setup Rule Files" to re-trigger setup
- Restart Cursor if needed
Rule files not installed
- Run
Cmd+Shift+P → "Agent Pipeline: Setup Rule Files"
- Check that
.cursor/rules/ contains global.mdc, planner.mdc, implementer.mdc, reviewer.mdc, test.mdc
Model not available in Cursor
- Make sure the model is enabled in your Cursor subscription
- You can type any model name in Settings — the dropdown list is just a convenience
License
MIT
