Intelligent Routing

If the domain is ambiguous, resolve that ambiguity before implementation.

Step 3 - Announce Agent

Always announce the selected agent before continuing:

🤖 Applying knowledge of @[agent-name]...

Do not skip this announcement.

Step 4 - Load Agent

Open .agent/agents/{agent}.md.

Then:

• read the agent rules
• read the frontmatter
• extract the skills: list
• apply the agent-specific constraints before proceeding

Step 5 - Load Skills

For each skill declared by the selected agent:

• open SKILL.md first
• read only the sections relevant to the current task
• do not load the entire skill blindly

Prefer minimal, targeted loading over broad context expansion.

Step 6 - Check Workflow

If the task matches a workflow, open .agent/workflows/<name>.md and follow it step by step.

Typical triggers include:

• planning flows
• creation flows
• debugging flows
• orchestration flows
• testing flows

Do not bypass a matching workflow.

Step 7 - Run Socratic Check

Before coding or design work, ask clarifying questions when:

• requirements are unclear
• the task is complex
• multiple files are likely to change

Never assume hidden requirements.

Step 8 - Execute

Only after classification, routing, agent loading, skill loading, workflow check, and Socratic check:

• generate code
• generate UI
• generate architecture
• generate the solution

Failure Conditions

Treat the following as protocol failures:

• no agent selected
• no agent announcement
• skills not loaded
• code or design created before routing
• workflow ignored

Success Condition

A valid response is:

• routed
• specialized
• consistent with the Antigravity system