Primary Orchestrator Agents
Relevant source files
The following files were used as context for generating this wiki page:
Primary Orchestrator Agents constitute the core intelligence and management layer of the project. These agents—primarily Sisyphus, Atlas, Prometheus, and Hephaestus—work in a hierarchical structure to parse user requests, generate strategic work plans, and execute tasks autonomously. They act as the "conductor" or "general" of the system, delegating specialized work to subagents and verifying outcomes to ensure high-quality software engineering output.
The architecture emphasizes a strict separation between planning and execution. While agents like Prometheus focus exclusively on foresight and requirements gathering, Sisyphus and Atlas coordinate the actual implementation phases, utilizing a delegation system based on categories and skills to route tasks to the most appropriate models.
Sources: src/agents/sisyphus.ts:152-168, src/agents/prometheus/identity-constraints.ts:13-35, AGENTS.md:144-156
Agent Hierarchy and Lifecycle
The workflow typically follows a structured sequence from intent analysis to final verification. The system uses a multi-agent loop to ensure that no implementation begins without a bulletproof plan.
Core Workflow Flowchart
The following diagram illustrates the standard progression from user request to completed work:
The workflow ensures that requirements are gathered (Metis), a plan is drafted (Prometheus), and the plan is reviewed (Momus) before any orchestrator (Atlas/Sisyphus) begins delegating to specialized workers.
Sources: src/hooks/prometheus-md-only/constants.ts:19-48, src/agents/prometheus/plan-generation.ts:16-30
Sisyphus: The Powerful Orchestrator
Sisyphus is the primary orchestrator, characterized as a "Senior SF Bay Area engineer." Its role is to parse implicit requirements and delegate specialized work to subagents.
Key Competencies
- Intent Gating: Classifies requests into types such as Trivial, Explicit, Exploratory, or Open-ended before acting.
- Delegation Bias: Sisyphus follows a "Delegate by Default" rule, working personally only on the most trivial tasks.
- Task Management: Mandates the creation of a
Todo List(orTask List) before starting any non-trivial work to ensure recovery and visibility.
Delegation Logic
Sisyphus utilizes a task() tool that combines Categories (domain-optimized models) and Skills (injected expertise).
| Category | Best For | Model |
|---|---|---|
visual-engineering | Frontend, UI/UX, styling | google/gemini-3-pro |
ultrabrain | Complex architecture, logic | openai/gpt-5.3-codex |
quick | Trivial fixes, single files | anthropic/claude-haiku-4-5 |
writing | Documentation, prose | google/gemini-3-flash |
Sources: src/agents/sisyphus.ts:152-168, src/agents/sisyphus.ts:258-272, src/agents/sisyphus.ts:312-325
Prometheus: The Strategic Planner
Prometheus is a dedicated planning agent that acts as a strategic consultant. It is strictly prohibited from writing code or executing implementation tasks.
Identity Constraints
Prometheus operates under a PLANNING_ONLY mandate. If a user asks to "fix a bug," Prometheus interprets this as "create a work plan to fix a bug."
The Work Plan Template
Prometheus generates markdown plans in .sisyphus/plans/*.md containing specific mandatory sections:
- TL;DR: Summary and deliverables.
- Context: Original request and interview summary.
- Verification Strategy: Mandatory agent-executed QA scenarios (no human intervention permitted).
- Parallel Execution Waves: Grouping independent tasks to maximize throughput.
- TODOs: Granular tasks with recommended agent profiles and reference citations.
Sources: src/agents/prometheus/plan-template.ts:10-50, src/agents/prometheus/plan-template.ts:158-180
Hephaestus: The Autonomous Deep Worker
Hephaestus is the "Senior Staff Engineer" agent optimized for repository-scale architecture and autonomous execution. Unlike other agents, it follows a "Keep Going" principle, attempting multiple creative approaches before asking the user for help.
Operational Principles
- Judicious Initiative: Hephaestus makes reasonable technical decisions without asking, covering all plausible intents when ambiguous.
- Exploration Hierarchy: It must try direct tools (git, grep, bash), then explore agents, then librarian agents, and finally context inference before reaching out to the user.
- Deep Reasoning: Uses GPT-5.3 Codex with medium reasoning effort for all modifications to maintain architectural consistency.
Completion Definition
Hephaestus only considers a task complete when:
lsp_diagnosticsreturns zero errors on all modified files.- The build command exits with code 0.
- All verification evidence is provided.
Sources: src/agents/hephaestus.ts:77-105, src/agents/hephaestus.ts:145-165, src/agents/hephaestus.ts:384-405
Atlas: The Master Orchestrator
Atlas serves as a high-level conductor, specifically optimized for long-running execution of Prometheus plans. It focuses on coordinating multiple waves of subagents.
Verification and QA
Atlas acts as the final QA gate. It does not trust subagent reports and must verify every delegation through:
lsp_diagnostics(".")for project-wide health.Bash("bun run build")for compilation status.Bash("bun test")for logic verification.
State Management via Notepads
Atlas maintains "cumulative intelligence" by using notepads in .sisyphus/notepads/{plan-name}/. It records learnings, decisions, and issues that are passed to subagents to ensure they aren't working in isolation.
| Notepad File | Purpose |
|---|---|
learnings.md | Successful patterns and discovered context |
decisions.md | Architectural choices made during execution |
issues.md | Blockers or regressions encountered |
Sources: src/agents/atlas/gpt.ts:11-25, src/agents/atlas/gpt.ts:124-140, src/agents/atlas/gpt.ts:187-200
Conclusion
The Primary Orchestrator Agents provide a robust framework for autonomous software development. By separating the strategic foresight of Prometheus from the deep execution capabilities of Hephaestus and the broad coordination of Sisyphus and Atlas, the system minimizes AI "slop" and ensures that code changes are verified, documented, and architecturally sound.