oh-my-ag - Orchestrate 6 specialized domain agents
In the recent *Claude 2026 Agentic Coding Trends Report*, software development is no longer defined as “writing code directly,” but as the challenge of how to orchestrate and supervise agents.
AI is a strong collaborator, but quality and security still require human oversight and validation. Full delegation may sound appealing, yet long-term maintenance demands a collaborative structure. The real issue is that this collaboration model usually has to be manually constructed and managed through prompts—every single time.
Why oh-my-ag
If you've used Antigravity or CLI-based agents in real projects, you’ve probably encountered situations like:
- Agents failing to properly read Skills
- Rules being ignored mid-conversation
- Responses falling into uncontrolled infinite loops
- Agents working hard—but not actually doing what was requested
At first glance, this looks like a model problem.
In reality, it was an orchestration problem.
Instead of endlessly refining prompts, we changed direction.
Rather than asking, “How do we write better prompts?”
We asked, “What if we provide agent collaboration as a default structure?”
That question led to oh-my-ag.
What oh-my-ag Does
oh-my-ag is a role-based agent orchestration layer for Antigravity.
Instead of pushing all context into a single agent, it clearly separates responsibilities.
- The orchestrator controls the overall flow
- Sub-agents focus on their specific roles
- Shared memory manages necessary context
Each agent is optimized with role-specific Skills and prompt structures, and the orchestrator binds them into a single coherent workflow.
Why Use It
1. Process mitigates model variability
Model performance can shift unexpectedly due to updates or infrastructure conditions. Architectures that rely on a single agent and long prompts are highly fragile.
oh-my-ag splits roles and responsibilities so that temporary degradation in one model does not compromise the entire result.
All agents share Serena Memory as a common memory layer. Decisions and intermediate outputs accumulate, reducing context loss even if models change or output quality fluctuates.
2. Clear human supervision points
The goal is not to remove humans from the loop, but to clarify where intervention is necessary.
By separating PM, QA, and Debug roles, it becomes structurally clear where AI operates and where human judgment should step in.
3. Reduced prompt management overhead
Rewriting the same role descriptions, rules, and approval flows in prompts is not productive.
oh-my-ag automates these defaults through role-based Skill bundles and a parallel-capable orchestrator.
Key Features
- Role-based specialized agent set
- Parallel-capable orchestration layer
- Supports Gemini CLI / Claude CLI / Codex CLI
- Serena Memory integration
- MCP-based tool scope control
- Conventional Commits automation
AI-native development in 2026 is no longer about “using tools well.”
This orchestration is designed around the first-fluke/fullstack-starter structure and already handles 50+ commits per day in full-stack web and mobile applications.
If you’re already using Antigravity, instead of wrestling with prompts, try running oh-my-ag with AI Pro—and burn tokens across six accounts at once.
GitHub: first-fluke/oh-my-ag
