Frontier Corp builds and deploys specialized AI engineering teams inside the repository. Frontier, formerly AgentX, is its Hypervelocity Engineering platform. Its fleet of Forward Deployed Engineers (FDEs) plans, designs, builds, reviews, tests, deploys, and captures what it learns while repo-local state and mechanical gates keep the work auditable.
Frontier combines specialized FDEs with the contracts and evidence required to ship dependable software.
Install | Feature list | See the workflow | Meet the fleet | Explore skills | Read the guide | Security
Frontier in 60 Seconds
| What you get | Current source |
|---|
| Frontier FDEs | 26 total: 15 visible roles and 11 internal specialists |
| Production knowledge | 134 skills across architecture, AI, data, development, design, testing, infrastructure, low-code, and domain consulting |
| Reusable delivery assets | 15 templates, 24 reusable prompt templates, and 7 root instruction files, plus scoped instruction sets |
| Quality discipline | Artifact-scoped evidenced iterations (1/2/3/5 after applicability), fresh verification, independent review, scrub, and completion gates |
| Durable artifacts | PRDs, ADRs, specs, UX prototypes, plans, reviews, learnings, issue state, and memory stored in the repository |
| Runtime surfaces | VS Code, GitHub Copilot Chat, GitHub Copilot CLI, Claude Code, Cursor, PowerShell, and Bash |
| Work tracking | Local mode, GitHub, or Azure DevOps |
| Model adapters | GitHub Copilot, Claude Subscription/API, OpenAI API, and Claude Code through LiteLLM + Ollama |
Source version and published package availability are separate. As checked on
2026-10-06, the Marketplace
serves extension 9.8.0, while GitHub releases
provide standalone assets through v9.6.0. The feature inventory
also includes Unreleased source improvements absent from the published 9.8.0 VSIX.
The core idea
Editable Mermaid source
Frontier Orchestration FDE can run that path in one session. For tighter control, select a specialist role for only the phase you need.
Execution depends on the host, configured provider, permissions, and available
tools. The GitHub workflow routes work and creates handoff scaffolds; it does not
itself prove delivery. Skill instructions and scripted demos are guidance, not
evidence that each advertised workflow passed a live end-to-end test. The current
AI evaluation gate tests issue classification, not the quality of all 26 roles.
Why Frontier
Repository context before generation
Agents retrieve project instructions, approved product and architecture artifacts, relevant skills, prior decisions, and known pitfalls before they write. The repository is the system of record, not the model's memory.
Evidence before completion
A passing-sounding response is not a gate. Frontier records test output, coverage, security scans, browser checks, artifact freshness, and independent review findings. HIGH and MEDIUM findings block handoff.
Role contracts instead of generic personas
Each role has a defined pipeline, deliverables, boundaries, templates, and done criteria. Product Manager produces requirements. Architect evaluates options. Engineer implements. Reviewer verifies. The roles do not collapse into one unconstrained prompt.
Learning that survives the chat window
Plans, progress, review findings, decisions, and promoted learnings remain in the workspace. Long-running work can compact or reset without losing the durable execution contract.
The Operating Loop
Frontier uses six shared checkpoints across chat, CLI, issues, plans, reviews, and VS Code sidebars:
| Checkpoint | Purpose | Durable evidence |
|---|
| Brainstorm | Frame the problem and retrieve prior learning | Issue or bounded task |
| Plan | Record scope, alternatives, risks, and verification | Execution plan and optional work contract |
| Work | Implement a bounded slice | Code, artifacts, progress, implementation evidence |
| Review | Inspect implementation, regression cases and non-test evidence | Structured findings, current checks and review decision |
| Compound Capture | Preserve reusable outcomes or record a skip rationale | Learning artifact or closeout rationale |
| Done | Close only when delivery and evidence agree | Completed owning loop when applicable, review, and capture state |
Quality gates
- Test suites run separately from local loops and reviews. After a completed
loop, Frontier asks whether to run the suite and waits for explicit approval.
Skipped suites remain not run; CI/release requirements are unchanged.
- Iterative loop: only create/update/delete/rename of PRD, TechSpec, code or unit tests requires a loop. Research/read/review-only/check/commit-only and other artifacts (plans, docs, config, instructions, declarations) are exempt; nonloop checks and council policy remain. Core protected executable code stays covered. Roles and keywords do not exempt covered work. After applicability, minimums scale by risk: standard
1, auto-fix 2, complex delivery 3, high-risk 5.
- Review scope: inspect the covered work's evidence without starting a review-only loop or inheriting unrelated active state. Covered fixes require the gate before editing; delegates reuse only the assigned parent loop and leave its lifecycle to the owner.
- Independent review: a subagent sees the deliverable, not the author's rationale
- Karpathy guidelines: think before coding, keep it simple, change surgically, verify the goal
- Model Council: Analyst, Strategist, and Skeptic pressure-test high-stakes decisions
- Deslop scrub: flags stale comments, over-abstraction, generic UI, and unsafe empty catches
- Skill quality: deterministic 100-point rubric with blocking floors and no-regression checks
- Fresh evidence: reused or stale artifacts cannot complete the loop
- Compound capture: reusable decisions and pitfalls become repository knowledge
The Frontier FDE Fleet
15 visible roles
| Role | Best used for | Primary output |
|---|
| Frontier Orchestration FDE | End-to-end autonomous delivery | Routed specialist workflow |
| Frontier Product FDE | Product scope and outcomes | PRD, roadmap, backlog |
| Frontier Experience FDE | User flows and accessible interfaces | UX spec and working prototype |
| Frontier Architecture FDE | Options, tradeoffs, and system boundaries | ADR and technical specification |
| Frontier Engineering FDE | Production implementation | Code, tests, evidence |
| Frontier Review FDE | Functional, security, and architecture review | Approval or actionable findings |
| Frontier Auto-Fix FDE | Review plus bounded safe fixes | Review and verified corrections |
| Frontier DevOps FDE | CI/CD and release automation | Pipelines, deployment, rollback |
| Frontier AI Systems FDE | Agent, RAG, prompt, and model quality | AI pipeline, eval plan, model card |
| Frontier Test FDE | Test strategy and release certification | Test suites and certification report |
| Frontier Fabric FDE | Microsoft Fabric data products | Lakehouse, Warehouse, Spark, pipelines, lineage |
| Frontier Power Platform FDE | Portable low-code solution source | Dataverse, apps, flows, PCF, Pages, Copilot Studio |
|
11 internal specialists
Frontier GitHub Ops FDE, Frontier ADO Ops FDE, Frontier ADO Planning FDE, Frontier Functional Review FDE, Frontier Architecture Review FDE, Frontier Prompt FDE, Frontier Evaluation FDE, Frontier Observability FDE, Frontier RAG FDE, Frontier Diagram FDE, and Frontier Prototype Audit FDE are invoked by parent roles when deeper isolation is useful.
134 Production Skills
Skills are compact, versioned knowledge packages that load only when the task needs them. Each SKILL.md can include scripts, references, and assets.
| Area | Examples |
|---|
| AI systems | Agent Framework, Foundry SDK, LangGraph, RAG, evaluation, safety, observability, memory, routing, voice agents |
| Architecture | API design, security, database, performance, cost analysis, infrastructure governance, low-code vs pro-code |
| Development | Testing, error handling, debugging, configuration, type safety, code review, prose anti-slop editing, worktrees, verification |
| Data and analytics | Fabric, Databricks, Cosmos DB, Power BI, forecasting, data analysis |
| Design | UX/UI, accessibility, prototype craft, anti-slop, content design, visual regression |
| Infrastructure | Azure, Bicep, Terraform, containers, GitHub Actions, YAML pipelines, release management |
| Low-code | Dataverse schema, canvas/model-driven apps, Power Automate, Power Pages, PCF, security roles, PAC CLI |
| Languages | C, C++, C#, Python, Go, Rust, React, Blazor, PostgreSQL, SQL Server |
| Consulting domains | Financial services, audit, tax, legal, oil and gas, CLM, corporate governance |
The executable skill gate scores specification, discoverability, decision support, actionability, safety, maintainability, and efficiency. Existing debt remains visible; changed skills cannot regress.
Platforms and Adapters
VS Code and GitHub Copilot
The Marketplace extension provides declarative chat agents, sidebars, Command Palette workflows, workspace initialization, adapter setup, and the bundled Frontier runtime.
LLM adapters
- GitHub Copilot for the default VS Code and CLI experience
- Claude Subscription through an authenticated local Claude Code CLI
- Claude API with workspace-scoped secret storage
- OpenAI API with workspace-scoped secret storage
- Claude Code + LiteLLM + Ollama for an Anthropic-compatible local gateway
Model names are advisory. Role boundaries, evidence requirements, and tool permissions remain the contract.
Work adapters
- Local: filesystem-backed issues and state for solo or offline work
- GitHub: issues, pull requests, Projects V2 status, and Actions
- Azure DevOps: work items and provider-aware delivery workflows
Editor and CLI portability
Frontier also ships GitHub Copilot CLI packs, Claude Code commands, Cursor rules/commands, and PowerShell/Bash launchers.
For Cursor, initialize the local runtime and run Frontier: Initialize Cursor.
Standalone installs use frontier cursor setup --restore-mcp. Setup preserves
user configuration and wires native context/policy hooks without copying agent
or skill trees. See the Cursor setup guide.
Selected Capabilities
Fabric Engineer
Builds Microsoft Fabric Lakehouse and Warehouse structures, OneLake shortcuts, Spark notebooks, data pipelines, medallion products, quality checks, reconciliation, lineage, and recovery documentation.
Power Platform Builder
Generates unpacked unmanaged solution source for Dataverse, canvas and model-driven apps, Power Automate cloud and desktop flows, Power Pages, PCF, plugins, security roles, environment variables, and Copilot Studio. The role is fail-closed for tenant mutation: it cannot authenticate, import, publish, export, or delete in a tenant.
Cost and infrastructure governance
Adds cost-envelope analysis, deterministic resource naming, and governance checks for Terraform, Bicep, and ARM. The scanner catches missing companion controls such as disabled authentication without identity and role assignment.
Secure WhatsApp companion
Controls a local Frontier workspace from an allowlisted account with read-only defaults, replay protection, short-lived sender-bound confirmation nonces, transcript-first voice handling, sandboxed Chromium, and bounded secret-reduced child execution.
Teams and GitHub App companion
View agent job progress and send confirmed instructions from Microsoft Teams or
GitHub issue/PR discussion comments. The collaboration companion
uses authenticated provider endpoints, user allowlists and a durable serialized
job queue. Follow-ups run as subsequent agent turns. Live use requires your bot
registration or GitHub App installation and an HTTPS endpoint.
Deterministic skill rubric
Replaces the legacy structural score with a 100-point rubric, strict YAML parsing, stable JSON evidence, blocking floors, and trusted-base no-regression enforcement for changed skills.
Security and Release Integrity
Frontier places controls outside the model prompt:
- destructive-command deny rules and surface-specific confirmation controls
- workspace path checks and secret redaction; shell commands are not OS-sandboxed
- SSRF validation with private-address and metadata-endpoint blocking
- role-specific tool boundaries, including fail-closed Power Platform terminal policy
- pinned GitHub Actions, dependency audits, secret scanning, and SAST
- release SBOMs and GitHub artifact provenance attestations
- fresh-evidence hashing for quality-loop iterations and completion
- WCAG, browser interaction, and anti-slop gates for UI-bearing work
See SECURITY.md for the threat model, supported versions, and reporting process.
Install in VS Code
1. Install the extension
code --install-extension jnPiyush.agentx
Or install Frontier from the Visual Studio Marketplace.
Requirements:
- VS Code 1.134.0+
- Git
- PowerShell 7.4+ (
pwsh) on every OS; the Bash launcher also delegates to PowerShell
- GitHub Copilot and GitHub Copilot Chat
2. Use Frontier in a trusted workspace
Open a trusted repository and use Frontier chat or commands. Private state is
created on first use; installation alone supplies the shared agents and skills.
For optional portable launchers and repository configuration, run:
Frontier: Initialize Repository Support
Or use chat:
@frontier initialize local runtime
Frontier uses a zero-copy runtime. Initialization creates local state, stable launchers, plans, reviews, and learning folders without copying the bundled agent/skill tree into your repository.
3. Add adapters when needed
Frontier: Add Remote Adapter
Frontier: Add LLM Adapter
Secrets are collected through secure VS Code prompts and stored in secret storage, not committed to .frontier/config.json.
Standalone workspace installers keep Frontier legal material under .frontier/runtime/legal and never replace the repository's root LICENSE or NOTICE. When upgrading an existing 9.x workspace installation to a different version, rerun with explicit -Force or --force so the installer cannot report a new version while retaining older managed files.
4. Start with one prompt
[Frontier Orchestration FDE selected]
Build a task tracker for small teams. Define the product, design the UX and architecture,
implement it, review the result, and capture reusable learning.
Or select a specialist and request one phase:
[Architect selected]
Evaluate three deployment options for this service and create an ADR with the tradeoffs.
Repository Map
| Path | Purpose |
|---|
| AGENTS.md | Workflow map, classifications, role pipelines, hard gates |
| Skills.md | Complete skill index and task-to-skill router |
| docs/WORKFLOW.md | Checkpoints, status transitions, handoffs, evidence model |
| docs/GUIDE.md | Setup, adapters, local mode, troubleshooting |
| .github/agents/ | Canonical agent role contracts |
| .github/skills/ | Canonical production skills |
| .frontier/runtime/ | CLI, MCP server, plugins, hooks and templates (runtime state lives in gitignored .frontier/) |
| docs/artifacts/ | PRDs, ADRs, specs, reviews, learnings |
docs/execution/ | Living plans, progress, contracts, evidence (created on use) |
| vscode-extension/ |
New In 9.8.0
Frontier 9.8.0 adds guided collaboration and zero-setup workspaces:
- clarify, propose a versioned plan and wait for explicit approval before
guided native execution; report milestones and resume durable sessions
- work in any folder without per-workspace initialization; private state is
created on first use and repository support stays optional
- retrieve bounded repository context from hierarchical summaries, AST symbols
and typed relationships with token budgets
- prepare quality loops faster with change-aware non-test preflight, reusable
receipts, factual review packets and phase timing
- continue guided sessions from the optional WhatsApp and Teams/GitHub companions
- revalidate approved reviews at commit and handoff, bound model and CLI calls,
and install compatible bundled plugins without a catalog download
AgentX/HVE compatibility aliases are removed; see
Frontier-only interfaces. Frontier
9.7.0 (repository graph, model routing and opt-in HydraFusion) is included.
Local validation and remaining provider/platform limitations are recorded with
the release package. Public publishing and experimental HydraFusion qualification
remain separate from packaging and source review.
Unreleased source improvements
- Stage learned skills for human review before explicit publication.
- Record post-loop test outcomes through the CLI without treating review approval
as a test pass.
- Append native model-usage metadata as responses arrive, and retain full final
clarification guidance without the former 600-character cut-off.
These additions are described in CHANGELOG.md;
an older installed VSIX does not acquire them merely because its version matches
the source version.
Read CHANGELOG.md for validation evidence, limitations, and prior releases.
Contributing
Start with CONTRIBUTING.md. Contributions should begin with an issue, keep changes reviewable, add or update tests, and preserve the quality-loop and evidence contracts.
License
Frontier is licensed under Apache License 2.0. Third-party notices are in NOTICE.
