oh-my-agent: The Multi-Agent Harness That Checks the Work

한국어 | 中文 | Português | 日本語 | Français | Español | Nederlands | Polski | Русский | Deutsch | Tiếng Việt | ภาษาไทย

Agents narrate success. oh-my-agent checks the artifacts.

Spawning parallel agents is the easy part. The hard part is knowing whether they actually did the work. "Tests pass, all criteria met" costs an agent nothing to say, and nothing inside that same session can contradict it.

oh-my-agent makes the claim falsifiable. A Stop hook refuses to end your session until your project's own typecheck / test / lint script exits 0. A gate command decides whether a workflow really ran by looking for the artifacts it must have left behind — and its JSON verdict, not the agent's summary, is the result. An independent judge with a fresh context re-verifies every criterion each round, including the ones that already passed. Every gate decision lands on an append-only event log you can read after the fact. Then it runs that same discipline across a dozen agent runtimes from one portable .agents/ directory.

Watch the full video (35s)

Quick Start

The install scripts below auto-install bun, uv, and serena if they're missing.

# macOS / Linux — auto-installs bun, uv & serena if missing
curl -fsSL https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.sh | bash

# Windows (PowerShell) — auto-installs bun, uv & serena if missing
irm https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.ps1 | iex

# Or manual (any OS, requires bun + uv + serena)
bunx oh-my-agent@latest

# All skills, deployed to every detected runtime
# (.claude, .cursor, .codex, .opencode, .github, .agents)
apm install first-fluke/oh-my-agent

# A single skill
apm install first-fluke/oh-my-agent/.agents/skills/oma-frontend

Pick a preset and you're ready:

Works With Every Agent

Verification is worth little if it's locked to one vendor. oh-my-agent keeps .agents/ as the single source of truth and projects it into each runtime's native layout, so every supported tool shares the same skills, workflows, rules, and gates — and switching vendors is a config change, not a migration.

Claude Code native + adapter	Codex CLI native + adapter	Antigravity native SSOT	Cursor native + adapter	Qwen Code native dispatch	Reasonix native-compatible
Pi native-compatible	OpenCode native-compatible	Amp native-compatible	GitHub Copilot symlinked skills	Grok Build native hooks	Kiro CLI native hooks + agents

_{& more}

Your Engineering Team

Instead of one AI doing everything (and getting confused halfway through), oh-my-agent splits work across specialized agents. Each one knows its domain deeply, has its own tools and checklists, and stays in its lane.

Beyond Code: Content & Research Pipelines

Separate from the engineering team, oma ships content and research pipelines built to the same engineering discipline: deterministic replay from fixtures, manifests for reproducibility, and honest degradation reporting when a source or vendor key is unavailable rather than a silently thinner result.

Orca IDE

OMA for Orca adds a sidebar panel and command-palette actions for project setup, review, debugging, verification, and local results. It uses your existing Orca agent terminal and project OMA installation. Requires Orca 1.4.197+ with experimental plugins enabled.

How It Works

Just chat. Describe what you want and oh-my-agent figures out which agents to use.

You: "Build a TODO app with user authentication"
→ PM plans the work
→ Backend builds auth API
→ Frontend builds React UI
→ DB designs schema
→ QA reviews everything
→ Done: coordinated, reviewed code

Or use slash commands for structured workflows:

Auto-detection: You don't even need slash commands — keywords like "architecture", "plan", "review", and "debug" in your message (in 11 languages!) auto-activate the right workflow. Detection accuracy is measured, not assumed: oma verify triggers scores the detector against a labeled 171-prompt corpus (currently 0% missed-fire, under 10% false-fire) and gates CI on it.

Per-Agent Models

Set model_preset in .agents/oma-config.yaml to choose which AI models each agent uses:

language: en
model_preset: mixed   # antigravity | claude | codex | cursor | kiro | mixed | qwen

# Optional per-agent overrides
agents:
  backend: { model: openai/gpt-5.5, effort: high }

• oma doctor --profile — prints the per-role resolved model matrix
• Full guide: web/docs/guide/per-agent-models.md

Verification, Not Narration

Each mechanism below is mechanical: a command exits 0 or it doesn't, a file is on disk or it isn't. No LLM is asked whether the work "looks correct."

Budgets are enforced the same way. session.quota_cap caps tokens, spawn count, and per-vendor spend; the orchestrator refuses the next spawn when a dimension is exceeded. When the wall-clock budget runs out, the Stop hook stops honestly with partial status recorded on the event log, rather than pretending completion.

Control Boundary

oh-my-agent leaves open-ended planning and next-action selection to the host LLM. It does not replace that judgment with a universal workflow graph or policy engine. Instead, it externalizes the invariants that must hold regardless of the model: tool guardrails, permissions, budgets, retry and stop limits, durable events, and mechanically verified completion. Structured events record decisions and gate outcomes; they do not act as a second planner.

Deterministic SLM execution is therefore a separate, optional product direction rather than missing infrastructure in the current harness.

Why oh-my-agent?

• Role-based — agents modeled like a real engineering team, not a pile of prompts
• Conditional context — dispatch loads the owning skill and defers supporting references until the task needs them. File-size scenarios and runtime loading are reported separately (measurement guidance).
• Recoverable — after 2 failed retries, orchestrate spawns hypothesis variants in parallel and keeps the highest-scoring result instead of retrying a wrong approach forever
• Monorepo-aware — detectWorkspace reads pnpm / nx / turbo / lerna and routes each agent to its workspace
• Multi-vendor — mix Antigravity, Claude, Codex, Cursor, Kiro, and Qwen per agent type
• Observable — terminal and web dashboards for real-time monitoring

Architecture

flowchart TD
    subgraph Workflows["Workflows"]
        direction TB
        W0["/brainstorm"]
        W1["/work"]
        W1b["/ultrawork"]
        W2["/orchestrate"]
        W3["/architecture"]
        W4["/plan"]
        W5["/review"]
        W6["/debug"]
        W7["/deepinit"]
        W8["/design"]
    end

    subgraph Orchestration["Orchestration"]
        direction TB
        PM[oma-pm]
        ORC[oma-orchestration]
    end

    subgraph Domain["Domain Agents"]
        direction TB
        ARC[oma-architecture]
        FE[oma-frontend]
        BE[oma-backend]
        DB[oma-db]
        MB[oma-mobile]
        DES[oma-design]
        TF[oma-tf-infra]
    end

    subgraph Quality["Quality"]
        direction TB
        QA[oma-qa]
        DBG[oma-debug]
    end

    Workflows --> Orchestration
    Orchestration --> Domain
    Domain --> Quality
    Quality --> SCM([oma-scm])

Learn More

• Detailed Documentation — Full technical spec and architecture
• Supported Agents — Agent support matrix across IDEs
• Capability Providers — Experimental Gortex and Honcho configuration, routing, and limits
• Benchmark Report — Method, scores, screenshots, and caveats
• Web Docs — Guides, tutorials, and CLI reference

Sponsors

This project is maintained thanks to our generous sponsors.

Like this project? Give it a star!

gh api --method PUT /user/starred/first-fluke/oh-my-agent

Try our optimized starter template: fullstack-starter

🚀 Champion

🛸 Booster

☕ Contributor

Become a sponsor →

See SPONSORS.md for a full list of supporters.

Star History

References

• Li, X., Liu, Y., Chen, W., You, B., Di, Z., He, Y., Zheng, S., Choe, K. W., Sun, J., Wang, S., Tao, C., Li, B., Zhao, X., Geng, H., Wu, X., Zhou, J., Chen, X., Xing, H., Li, Y., … Song, D. (2026). SkillsBench: Benchmarking how well agent skills work across diverse tasks (Version 4) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2602.12670
• Yu, G., & Wang, X. (2026). Knows: Agent-native structured research representations (Version 1) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2604.17309
• Liang, Q., Wang, H., Liang, Z., & Liu, Y. (2026). From skill text to skill structure: The scheduling-structural-logical representation for agent skills (Version 4) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2604.24026
• Chen, C., Yu, Q., Gu, Y., Huang, Z., Li, H., Liu, H., Liu, S., Liu, J., Peng, D., Wang, J., Yan, Z., Meng, F., Qin, E., Che, C., & Hu, M. (2026). The scaling laws of skills in LLM agent systems (Version 1) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.16508
• Tang, L., Rashtchian, C., Ferng, C.-S., Tomkins, A., Juan, D.-C., & Vu, T. (2026). WikiSkill: Compiling agent experience into persistent knowledge for skill evolution [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2608.27454
• Huang, Z., Xu, J., Yang, Y., Gong, Z., Yang, Q., Tian, M., Wang, X., Lv, C., Gao, X., Dai, Q., Liu, B., Qiu, K., Yang, X., Chen, D., Zheng, X., & Luo, C. (2026). From raw experience to skill consumption: A systematic study of model-generated agent skills [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.23899
• Hong, D. B., Imani, A., & Ahmed, I. (2026). From anatomy to smells: An empirical study of SKILL.md in agent skills (Version 2) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2607.01456

License

MIT