senior-agentic-enginner

Create a skill when the requirement is specialized knowledge or a repeatable procedure that an existing agent can load. Create a new agent only when the responsibility cannot reasonably belong to an existing agent, since agents represent stable responsibility boundaries while skills represent reusable capabilities.

Treat those files as input material, not final architecture. Inventory and read them, classify each instruction into agent responsibility, skill knowledge, workflow procedure, project context, or platform adapter, then move each piece to its canonical location and build an index.md mapping layer.

Yes, the canonical architecture of agents, skills, workflows, and context is platform-neutral. Claude Code, Codex, Hermes, and OpenCode are treated as runtime adapters that map their own discovery and execution concepts onto the project's canonical system.

Load skills just in time rather than attaching every skill to every agent, and use compact handoff artifacts between agents instead of passing entire conversation history. Each agent should receive only the minimum context needed for its next correct decision.

A review detects responsibility duplication between agents, overlapping skills, duplicate workflows, circular dependencies, excessive context loading, platform coupling, conflicting instructions, hidden dependencies, and monolithic agents accumulating delegated responsibilities.