Agents
Helm uses 13 agents — 11 pipeline agents, 1 on-demand, and 1 orchestrator — to guide projects from discovery to deployment. Each agent has a defined mission, inputs, outputs, and success criteria.
Quick Reference
| Agent | Phase | Question | Threshold |
|---|---|---|---|
| Scout | Discover | What are we building? | 85% |
| Survey | Discover | How does it work today? | 85% |
| Research | Discover | What is the problem? | 90% |
| Planning | Plan | What exactly will we build? | 90% |
| Architect | Plan | How will we build it? | 90% |
| Roadmap | Plan | When will we build it? | 90% |
| Breakdown | Plan | Who does what specifically? | 90% |
| Review | Quality | Is everything traceable? | 95% |
| Build | Build | Build it. | 90% |
| Verify | Quality | Does it work correctly? | 95% |
| Ship | Deploy | Ship it. | 90% |
Discover Phase
Scout (Greenfield)
Mission: Extract project vision, goals, and constraints through guided conversation.
Scout is the first agent for new projects. It asks structured questions about what you want to build, who the users are, what constraints exist, and what success looks like. Output: operational context document.
Survey (Brownfield)
Mission: Comprehensive analysis of an existing codebase — tech stack, architecture, integrations, and technical debt.
Survey runs automated scans before asking questions. It maps folder structure, identifies patterns, categorizes debt by severity, and captures operational context. Output: survey report with debt inventory.
Research
Mission: Extract, analyze, and document core problems, desired outcomes, and constraints.
Research is the most critical discovery agent. It guides you through brain dump extraction, structured analysis, gap identification, and synthesis. Every downstream agent traces back to what Research captures. Output: research report.
Plan Phase
Planning
Mission: Transform research findings into a formal Product Requirements Document (PRD).
Every requirement has an ID, description, MoSCoW priority, research reference, and Given-When-Then acceptance criteria. Output: PRD with functional and non-functional requirements.
Architect
Mission: Design the technical architecture that fulfills PRD requirements.
Defines tech stack, system components, data model, API design, security model, and deployment strategy. For brownfield projects, assesses existing stack against new requirements. Output: architecture document.
Roadmap
Mission: Break PRD requirements into ordered development phases.
Phase 1 is always the MVP. Maps dependencies, defines deliverables per phase, and creates execution order. Output: phases document.
Breakdown
Mission: Decompose phases into atomic, testable tasks.
Each task has an ID, title, description, requirement reference, size estimate (XS to L), dependencies, and Given-When-Then acceptance criteria. No task exceeds 4-8 hours. Output: task breakdown.
Quality Phase
Review (Plan Gate)
Mission: Validate traceability across all planning artifacts.
Checks 4 chains: Research → PRD, PRD → Phases, Phases → Tasks, Tasks → Acceptance Criteria. Each orphaned requirement incurs a penalty. Requires 95% to proceed to Build. Output: review report with verdict.
Verify (Build Gate)
Mission: Validate implementation against acceptance criteria.
Runs tests, SAST scanning, code review against architecture patterns, and acceptance criteria verification. Requires 95% to proceed to Deploy. Output: verification report with verdict.
Build Phase
Build
Mission: Implement each task from the approved breakdown.
Works through tasks in dependency order. Self-validates against acceptance criteria. Can operate in interactive (confirm each task) or autonomous mode (auto-proceed when gates pass). Output: implemented code + build report.
Deploy Phase
Ship
Mission: Handle git operations, PR creation, and deployment.
Creates commits, pushes to remote, opens pull requests, and manages deployment. Output: deployment summary.
On-Demand
Tradeoff
Mission: Analyze multi-option decisions with structured pros/cons.
Can be invoked at any point by any agent when facing a decision with multiple viable approaches. After analysis, control returns to the invoking agent.