OmO
Oh My OpenAgentv3.17.13
AgentsAgents

Agents

The OmO agent catalog and responsibility split.

Core Ordering

The core agents have deterministic tab order in OpenCode: Sisyphus, Hephaestus, Prometheus, then Atlas. Other agents follow after that stable core.

AgentRoleNotes
SisyphusMain orchestrator. Reads intent, delegates, tracks todos, and finishes.Best with Claude Opus/Kimi/GLM style orchestration, with a GPT-5.4 path available.
HephaestusGPT-native autonomous deep worker.Use for architecture, deep debugging, or cross-domain synthesis.
PrometheusStrategic planner and interviewer.Creates plans; should not edit production code.
AtlasExecutes approved plans.Reads Prometheus plans, creates boulder state, delegates units, verifies progress.
OracleRead-only architecture/debugging consultant.Good for unfamiliar patterns, security/performance tradeoffs, hard debugging.
LibrarianDocumentation, remote repositories, OSS examples.Keep fast and cheap; retrieval should not use orchestration-grade tokens.
ExploreFast codebase search.Best for “where is X?” and pattern discovery.
Multimodal-LookerImages, PDFs, diagrams, screenshots.Uses visual-capable models for extraction and summary.
MetisPre-plan gap analyzer.Finds hidden intentions, ambiguity, missing constraints.
MomusPlan/result reviewer.Rejects vague or unverifiable plans.
Sisyphus-JuniorCategory-spawned executor.Focused worker; cannot re-delegate.

Tool Boundaries

AgentConstraint
OracleRead-only consultation. No writes, edits, or delegation.
LibrarianResearch-only. No writes, edits, or delegation.
ExploreSearch-only. No writes, edits, or delegation.
Multimodal-LookerMedia analysis, read-focused.
AtlasOrchestrates plan execution but delegates implementation units.
Sisyphus-JuniorImplements one assigned unit and must verify before returning.

These restrictions are part of the harness. They prevent “smart” agents from doing unsafe work in the wrong role.

Category Workers

When the task tool receives a category, OmO creates a Sisyphus-Junior-like worker with the category's model/profile. The category describes the domain; skills add instructions and MCPs.

CategoryWorker behavior
visual-engineeringFrontend/UI/design work. Pair with frontend-ui-ux and browser verification.
ultrabrainHard reasoning and architecture. Keep the prompt goal-oriented.
deepAutonomous research plus implementation. Good for hairy but coherent work.
artistryCreative or non-conventional approaches.
quickSmall changes. Do not use for visual work just because the diff is small.
writingDocumentation and prose.

Background Agents

Background agents let the main session keep working while specialists search or analyze.

task({
  subagent_type: "explore",
  load_skills: [],
  run_in_background: true,
  prompt: "Find all auth policy implementations and summarize file paths.",
});

Use background mode for independent research. Use synchronous mode when the result blocks the next implementation decision.

On this page