AgentX for VS Code
The IDE Orchestrator for Multi-Agent Software Delivery
Bring structured multi-agent workflows directly into your editor with chat execution, live workspace state, and seamless repo integration.
Why Use the Extension?
Running autonomous agents from the CLI lacks visibility. The AgentX VS Code extension bridges the gap, allowing you to trigger complex delivery pipelines while retaining absolute visibility and control over what the agents are thinking, validating, and writing.
"Full autonomous orchestration, deeply integrated with your local workspace."
The Extension Surface
| Feature |
Description |
| 26 Declarative Chat Agents |
15 visible lifecycle agents, including AgentX Auto for end-to-end orchestration, plus 11 hidden specialists that remain parent-invocable. |
| Model Council (core) |
Multi-model deliberation on high-stakes decisions -- Analyst + Strategist + Skeptic debate PRD scope, ADR options, AI design, code reviews, and deep research before they ship. Agent-internal by default; optional gh models multi-vendor auto-invoke. Mandatory gate for PM, Architect, Reviewer, Data Scientist, and Consulting Research on high-stakes work. |
| Copilot Chat Participant |
Native @agentx chat participant for triggering routines, brainstorm, learnings, and compound-loop inspection. |
| Karpathy Guidelines (built-in) |
The karpathy-guidelines skill is auto-loaded for Engineer, Architect, Reviewer, Auto-Fix Reviewer, DevOps, Tester, and Data Scientist -- enforcing think before coding, surgical diffs, assumption audits, and goal-driven execution to block common LLM coding pitfalls at authoring and review time. |
| Workspace Setup Wizard |
Local-runtime-first setup with optional remote adapters for GitHub or Azure DevOps and configurable LLM adapters. |
| 4 Sidebar Views |
Work (queues, workflow next step, brainstorm, learnings), Status (agent states, loop, dependencies, evaluation), Templates (output templates), Skills (134 production skills). |
| 50 Commands |
Workflow, loop management, knowledge compounding, AI evaluation, task bundles, bounded parallel delivery, and plugin management from the Command Palette. |
| Knowledge Compounding |
Ranked learnings, compound-loop visibility, learning-capture scaffolds, durable review-finding promotion, and agent-native review parity checks. |
| AI Evaluation |
Scaffold, run, and inspect AI evaluation contracts with rubric-based quality gates. |
| Task Bundles & Bounded Parallel |
Create, resolve, and promote scoped task bundles; run and reconcile bounded parallel delivery slices. |
| Plugin System |
Extend the workspace with Add Skill, Add Agent, and Add Plugin commands. |
Architecture Flow
flowchart LR
Chat["Copilot Chat"] --> Context["AgentX Context"] --> Engine["Execution Engine"]
Engine --> View["Sidebar and Live State"]
Engine --> File["Repo-Defined Assets"]
View -.->|"Queues and Workflows"| UI["VS Code UI"]
File -.->|"Skills and Templates"| Workspace["Local Files"]
- Inputs: VS Code Chat drives intent into the orchestrator.
- Control: The IDE tracks progress and state live via dedicated UI extensions.
- Outputs: Everything resolves natively into your repository as standard Markdown tracking, code, and CI manifests.
Requirements
To run AgentX successfully within VS Code:
- VS Code: 1.85.0 or newer
- System: Git configured on your PATH
- Runtime: PowerShell 7.4+ (
pwsh) on Windows, or Bash on Linux/macOS
- Integrations: gh (GitHub CLI) optional for extended GitHub mode operations
Quick Start
- Install the extension from the VS Code Marketplace.
- Open your target project workspace in VS Code.
- Initialize the workspace by running
AgentX: Initialize Local Runtime from the Command Palette, or start the same flow in chat with @agentx initialize local runtime.
- Optionally add a remote adapter with
AgentX: Add Remote Adapter or start it in chat with @agentx connect github, @agentx connect ado, @agentx use local, or @agentx add remote adapter.
- Optionally switch the workspace LLM adapter with
AgentX: Add LLM Adapter or start it in chat with @agentx switch llm, @agentx connect claude, @agentx connect claude local, @agentx connect openai, or @agentx use copilot.
- Select a role in Copilot Chat and run the next step for that role, or select AgentX Auto to orchestrate the full flow in one session.
- Capture reusable outcomes with
AgentX: Create Learning Capture once review confirms the result should compound future work.
Workspace Initialization
AgentX initialization is workspace-scoped. After opening a repository or project folder in VS Code, run:
AgentX: Initialize Local Runtime
You can also start the same flow in chat with:
@agentx initialize local runtime
This prepares the local AgentX runtime for the current workspace by:
- creating local runtime folders and state files
- preparing repo-local execution artifacts such as plans, progress, reviews, and learnings
- writing stable
.agentx/* workspace entrypoints that delegate into the bundled runtime
- keeping the executable runtime bundled while workspace state stays local to the repo
Repeat this step for each workspace where you want AgentX to run.
Optional Remote Integration
If you want GitHub or Azure DevOps issue and workflow operations, run:
AgentX: Add Remote Adapter
You can also start repo-adapter setup in chat with:
@agentx add remote adapter@agentx connect github@agentx connect ado@agentx use local
The extension now keeps repo-adapter setup conversational. Non-secret values are collected in chat, pending setup survives between turns, and the chat UI offers follow-up actions to continue or cancel the flow.
Stay on local runtime only when you want repo-local planning, implementation, and review without remote backlog integration.
Workspace LLM Adapter Setup
If you want to switch the workspace away from the default Copilot-backed path, run:
AgentX: Add LLM Adapter
You can also start LLM setup in chat with:
@agentx switch llm@agentx connect claude@agentx connect claude local@agentx connect openai@agentx use copilot
The extension now keeps LLM setup conversational. Non-secret values are collected in chat, pending setup survives between turns, and secret-bearing steps use VS Code's secure password prompt instead of asking you to paste keys into the chat transcript.
Available workspace LLM adapters include GitHub Copilot, Claude Subscription, Claude Code + LiteLLM + Ollama, Claude API, and OpenAI API. The local Claude option keeps claude-code as the execution transport while injecting Anthropic-compatible LiteLLM gateway settings and pinning the runner to the configured local coding model.
Use AgentX in the Agents Window
VS Code's Agents Window (Preview) lets supported chat participants run as first-class agents. AgentX opts in per user, not per workspace, because the underlying VS Code setting (extensions.supportAgentsWindow) lives in your user settings.json.
You have three ways to enable it:
Automatic prompt (recommended). The first time you install AgentX -- and again after each major-version upgrade -- the extension shows a one-time information message offering to enable AgentX in the Agents Window. Choose Enable in Agents Window, then reload the window when prompted. Choose Don't ask again to silence the prompt permanently.
Manual command. Run AgentX: Enable in Agents Window from the Command Palette at any time. The command is idempotent and preserves any other entries already in extensions.supportAgentsWindow.
Power users. Add the following to your user settings.json directly:
"extensions.supportAgentsWindow": {
"jnPiyush.agentx": true
}
After enabling, reload the VS Code window. AgentX will appear in the Agents Window agent picker alongside other opted-in extensions. To opt back out, remove the jnPiyush.agentx entry (or set it to false) in user settings.json and reload.
Build Software With AgentX
Once a workspace is initialized, you can use AgentX inside VS Code to move an app from planning through review.
flowchart LR
I[Install Extension] --> W[Open Workspace]
W --> R[Initialize Local Runtime]
R --> B[Select Role Or AgentX Auto]
B --> E[Create Or Execute Work]
E --> V[Review And Validate]
V --> C[Capture Learnings]
Recommended Flow
In VS Code, select the role in chat first, then send a prompt for that role. For example, if you are building a simple task-tracker app for small teams:
| Step |
Role |
What To Do |
Sample Prompt |
| 1. Define the product |
Product Manager |
Create the product scope, goals, and acceptance criteria |
Create a PRD for a task-tracker app for small teams with email login, task CRUD, due dates, and a dashboard for overdue work. |
| 2. Shape the UX |
UX Designer |
Turn the PRD into user flows and prototype-ready screens |
Create the user flow and prototype plan for the task-tracker app, covering sign-in, task creation, task filtering, and dashboard views. |
| 3. Design the architecture |
Architect |
Define the technical approach and key tradeoffs |
Create an ADR and tech spec for the task-tracker app using a web frontend, backend API, persistence, and role-based access. |
| 4. Implement the app |
Engineer |
Build the code and tests from the approved artifacts |
Implement the task-tracker app from the PRD and spec, including authentication, task CRUD APIs, dashboard data, and automated tests. |
| 5. Review the result |
Reviewer |
Check correctness, risk, and missing coverage before sign-off |
Review the task-tracker implementation for correctness, security, regressions, and missing tests. |
| 6. Preserve the learning |
AgentX Auto |
Capture reusable guidance from the work you just completed |
Create a learning capture for the task-tracker delivery workflow and major implementation lessons. |
If you want one orchestrated session instead of switching roles manually, select AgentX Auto and use one prompt such as:
Build a task-tracker app for small teams. Start by creating the PRD, then produce UX and architecture guidance, implement the app, review it, and capture reusable learnings.
Typical Chat Prompts
[Product Manager selected] Create a PRD for a task-tracker app for small teams
[UX Designer selected] Create the primary flows and screen plan for the task-tracker app
[Architect selected] Create an ADR and implementation spec for the task-tracker app
[Engineer selected] Implement the task-tracker app and its tests from the approved artifacts
[Reviewer selected] Review the task-tracker app implementation before sign-off
[AgentX Auto selected] Create a learning capture
When To Use Which Mode
- Use AgentX Auto when you want end-to-end orchestration in one session.
- Use a specialist role such as Product Manager, Architect, Engineer, or Reviewer when you want tighter control over one phase.
- Use the Command Palette and sidebars when you want a more guided workflow inside VS Code.
Compound Loop In The IDE
AgentX exposes the compound-engineering loop directly in VS Code instead of leaving it implicit in docs alone.
Chat Entry Points
@agentx brainstorm <topic> to start planning from ranked prior learnings@agentx learnings planning and @agentx learnings review <topic> to inspect curated guidance@agentx compound to view the current compound loop state@agentx create learning capture to scaffold a durable learning artifact for the active issue context@agentx review findings and @agentx agent-native review to inspect review-time follow-up surfaces
- Work sidebar:
Brainstorm, Planning learnings, Review learnings, Compound loop, Create learning capture
- Status sidebar:
Compound loop, Create learning capture, Agent-native review, Review findings, AI Evaluation Status
- Command palette equivalents exist for each of the same surfaces under the
AgentX: prefix
| View |
Contents |
| Work |
Workflow next step, brainstorm guidance, planning and review learnings, compound loop, learning capture, ready queue, and workflow rollout surfaces. |
| Status |
Agent status, loop state, dependency checks, AI evaluation, review findings, task bundles, bounded parallel runs, and digests. |
| Templates |
All output templates (PRD, ADR, Spec, UX, Review, Security Plan, Progress, Roadmap, Exec Plan, Contract, Evidence). |
| Skills |
134 production skills across 14 categories (architecture, development, languages, operations, infrastructure, data, documents, AI systems, design, testing, domain, product, diagrams, low-code). |
Command Reference
Workspace Setup
| Command |
Description |
| Initialize Local Runtime |
Prepare local runtime for the current workspace |
| Enable in Agents Window |
Opt AgentX into the VS Code Agents Window (Preview) for the current user |
| Add Remote Adapter |
Connect GitHub or Azure DevOps for backlog integration |
| Add LLM Adapter |
Switch the workspace LLM adapter (Copilot, Claude, OpenAI) |
| Add Plugin |
Extend the workspace with additional capabilities |
| Add Skill |
Add a production skill to the workspace |
| Add Agent |
Add an agent definition to the workspace |
Workflow
| Command |
Description |
| Show Workflow Next Step |
Show the recommended next action based on current checkpoint |
| Deepen Plan |
Refine the current execution plan |
| Kick Off Review |
Initiate the review phase |
| Show Workflow Steps |
Display the full workflow step list for a role |
| Show Workflow Rollout Scorecard |
View rollout readiness scores |
| Show Operator Enablement Checklist |
View the operator enablement checklist |
Quality Loop
| Command |
Description |
| Loop: Start |
Start a new quality loop iteration |
| Loop: Status |
Check current loop state |
| Loop: Iterate |
Record a loop iteration pass |
| Loop: Complete |
Mark the loop as complete |
| Loop: Cancel |
Cancel the active loop |
| Iterative Loop |
Run the full iterative loop flow |
Knowledge Compounding
| Command |
Description |
| Show Brainstorm Guide |
Start planning with ranked prior learnings |
| Show Planning Learnings |
View ranked planning learnings |
| Show Review Learnings |
View ranked review learnings |
| Show Knowledge Capture Guidance |
View capture guidance for the current context |
| Show Compound Loop |
Inspect the compound-engineering loop state |
| Create Learning Capture |
Scaffold a durable learning artifact |
| Show Agent-Native Review |
Run advisory agent-native review parity checks |
| Show Review Findings |
Inspect durable review findings |
| Promote Review Finding |
Promote a finding into a standard AgentX issue |
AI Evaluation
| Command |
Description |
| Show AI Evaluation Status |
View AI evaluation contract and results |
| Scaffold AI Evaluation Contract |
Create a new evaluation contract |
| Run AI Evaluation |
Execute an evaluation run |
Task Bundles & Parallel Delivery
| Command |
Description |
| Show Task Bundles |
View scoped task bundles |
| Create Task Bundle |
Create a new task bundle |
| Resolve Task Bundle |
Mark a task bundle as resolved |
| Promote Task Bundle |
Promote a bundle to an issue |
| Show Bounded Parallel Runs |
View active parallel delivery runs |
| Assess Bounded Parallel Delivery |
Assess readiness for parallel delivery |
| Start Bounded Parallel Delivery |
Launch a bounded parallel delivery slice |
| Reconcile Bounded Parallel Run |
Reconcile a completed parallel run |
Status & Diagnostics
| Command |
Description |
| Show Agent Status |
View agent states and active work |
| Check Dependencies |
Check issue dependency blockers |
| Generate Weekly Digest |
Generate a weekly status digest |
| Show Issue Detail |
View detailed issue information |
| Show Pending Clarification |
Check for pending clarification requests |
| Check Environment |
Validate the AgentX runtime environment |
| Refresh |
Refresh all sidebar views |
Chat Agents
The extension registers 26 declarative chat agents: 15 visible lifecycle agents
listed below and 11 hidden specialists that remain parent-invocable.
| Agent |
Role |
Use For |
| AgentX Auto |
Autonomous orchestrator |
End-to-end delivery in one session |
| Product Manager |
PRD and backlog |
Product scope, goals, stories |
| UX Designer |
UX and prototypes |
User flows, wireframes, HTML/CSS prototypes |
| Architect |
Architecture |
ADR, tech spec, tradeoff analysis |
| Engineer |
Implementation |
Code, tests, quality loop |
| Reviewer |
Code review |
Review, findings, approve/reject |
| Auto-Fix Reviewer |
Review + fix |
Review with safe auto-applied fixes |
| DevOps |
CI/CD |
Pipelines, deployment automation |
| Data Scientist |
ML/AI |
ML pipelines, evaluation, drift |
| Tester |
Testing |
Test suites, certification |
| Fabric Engineer |
Data platform |
Fabric Lakehouse, Warehouse, notebooks, pipelines, data quality |
| Power Platform Builder |
Low-code solutions |
Dataverse, apps, flows, Pages, PCF, Copilot Studio source |
| Power BI Analyst |
Reports |
Power BI, DAX, semantic models |
| Consulting Research |
Research |
Domain research, client materials |
| Agile Coach |
Stories |
Story creation, INVEST refinement |
Recent Changes
9.0.0
- Quality-loop approval now requires an attributable structured reviewer verdict with zero HIGH/MEDIUM findings on the final work iteration.
- Commit-time gates enforce risk-based
1/2/3/5 iteration minimums, staged/worktree agreement, and post-commit loop consumption.
- Autonomous workspace tools reject traversal, alternate streams, credentials, protected gate paths, links, aliases, and hardlinks.
- Autonomous shell execution and Claude-native tools remain disabled until an externally sandboxed adapter is available.
- Regression suites cover review exhaustion, hook lifecycle, path controls, staged and untracked harness enforcement, and VS Code evidence forwarding.
8.7.1
- Hardened fixed-source release recovery with tag, release-target, source-version, master-reachability, and checkout-SHA validation before repository scripts execute.
- Added SBOMs, SLSA provenance, and recovery-source attestations to recovered VSIX and MCP artifacts.
- Required Marketplace publication to verify provenance and the exact embedded publisher, extension name, and version while isolating the publish-only PAT to the final upload step.
- Fixed clean release packaging by installing extension dependencies before bundled asset synchronization.
- Made stamped-version release detection work for both linear and merge commits.
- Made version stamping portable across LF and CRLF package locks.
8.7.0
- Migrated agent defaults and provider routing to Claude Opus 5 and Sonnet 5.
- Added cost optimization and infrastructure governance skills with supply-chain, SSRF, and evaluation hardening.
- Added Fabric Engineer and promoted Power Platform Builder into core AgentX, bringing the inventory to 26 agents (15 visible, 11 internal).
- Added fail-closed Power Platform terminal enforcement, domain routing, canonical handoffs, installer parity, and adversarial regression coverage.
- Hardened the local WhatsApp companion with read-only defaults, confirmation-gated mutation, replay and voice safeguards, bounded CLI execution, sandboxed Chromium, and zero production audit findings.
- Added a deterministic 100-point skill-quality rubric with strict YAML, stable JSON evidence, blocking floors, trusted-base changed-skill enforcement, and Windows/POSIX installer parity.
- Release validation passed at the time of that release: extension coverage and 1013 tests, WhatsApp 23/23 with 90%+ line coverage, skill rubric behavior and 130-skill inventory validation, frontmatter 623/623, and zero HIGH/CRITICAL production dependency findings.
8.4.68
- Claude-backed AgentX defaults now use Claude Opus 4.8 across runtime model maps, VS Code adapter setup, agent creation pickers, and bundled agent definitions.
- Workspace-local launchers now keep loop state in their own workspace even when
AGENTX_WORKSPACE_ROOT leaks from another process, while extension-bundled runtimes still support explicit workspace roots.
- Bundled AgentX assets were regenerated for 8.4.68, including pack manifests, installers, docs, skills, and runtime scripts.
- Release validation passed: extension tests 913 passing, provider behavior 97/97, framework self-tests 134/134, and runner behavior 163/163.
8.4.63
- Model Council deepened into persona+purpose-specific deliberation (PRD scope, ADR options, AI design, code review, research) with multi-topic support in a single run
- Council persona model defaults refreshed to the current frontier tier (Opus 4.7 -> 4.8, GPT 5.4 -> 5.5); model names remain advisory diversity slots
- Extension opts into the VS Code Agents Window on activation as a user-side setting (SPEC-400) so AgentX surfaces in the agent-first window without leaving the editor experience
- Runtime hardening: resolved review-400 findings, restored quality-loop parity, and fixed a shell test flake
8.4.52
- New
convert-slides plugin renders Markdown storyboards into Microsoft PowerPoint (.pptx) via Pandoc, alongside the existing convert-docs (MD->DOCX) plugin
- AgentX Auto agent documents both plugins with trigger conditions and invocation rules (PATH precheck, no shell concatenation, regenerate-from-Markdown discipline)
- Consulting Research agent adopts a Markdown-first plugin-bridge workflow: storyboard Markdown is the source of truth and is rendered to
.pptx only on explicit request
- Zero-copy asset rewrite regression fix: agent context loader, runtime asset utilities, and agent-native review surface correctly resolve canonical template references through the bundled extension path (16/16 tests green)
8.4.51
- Bump version, sync bundled extension assets, repackage VSIX
8.4.47 - 8.4.49
- Bundled-asset sync fixes and VSIX repackaging across point releases
8.4.39
- MCP-only Azure DevOps provider: ADO work-item operations route through the official
@azure-devops/mcp server with configurable tool overrides
8.4.36
- New
product/prd skill: PRD authoring available to non-PM agents (Engineer, Architect, Auto) with a requirements-quality catalogue, vague-vs-concrete examples, and an AI-contract worked example
- New
diagrams/diagram-as-code skill: Mermaid, PlantUML, C4/Structurizr, Graphviz, and draw.io patterns with first-class support for cross-functional swimlanes, BPMN, and Visio (.vsdx) interop
- New internal
diagram-specialist sub-agent wired into Architect, Engineer, PM, UX Designer, Data Scientist, Reviewer, and Power BI Analyst
8.4.35
- Model Council mechanism: opt-in multi-perspective brief (Analyst, Strategist, Skeptic) for PRD scope, ADR options, AI design, code review, and research, completed agent-internally without involving the user
- New
karpathy-guidelines skill wired into Engineer, Architect, Reviewer, Auto-Fix Reviewer, DevOps, Tester, and Data Scientist to reduce common LLM coding pitfalls
8.4.30
- Updated agent model assignments across core roles (AgentX Auto, PM, Architect, Engineer, Reviewer, Auto-Fix Reviewer)
8.4.29
- Fixed ADO provider
--project flag handling for work item operations
- Provider-aware issue counting in the Work sidebar
- Closed stale issues with evidence-backed comments
8.4.28
- Bounded parallel delivery: assess, start, and reconcile parallel work slices
- Task bundle create, resolve, and promote commands
- Plugin system with
Add Plugin, Add Skill, Add Agent
- AI evaluation contract scaffolding and execution
8.4.25
- Workspace LLM adapter setup (Claude, OpenAI, Claude Code + LiteLLM)
- Conversational repo-adapter setup with pending state across turns
- Secure secret collection via VS Code password prompt
Earlier
- Compound loop, brainstorm, and knowledge-capture surfaces (8.4.7)
- Workspace initialization and remote adapter setup (8.4.0)
- Full sidebar views for Work, Status, Templates, Skills
Learn More
| |
