Feature Request: Automatic agent routing based on task context
Summary
When building multi-agent workflows with Claude Code, users currently need to define routing rules manually in CLAUDE.md and explicitly orchestrate which subagent handles which task. This works, but requires upfront configuration and Claude must follow hand-written rules.
Use Case
We use Claude Code with specialized subagents for different task types (analysis, implementation, security review, documentation). Our CLAUDE.md defines routing like:
| Task | Agent |
|---|---|
| Analysis, exploration | talos-analyst |
| Code writing, bugfix | talos-impl |
| Security review | talos-security-reviewer |
| Docs, version, changelog | talos-doc-updater |
This works well — but it requires manually maintaining routing rules in CLAUDE.md and trusting Claude to follow them consistently.
Feature Request
A built-in automatic context router that selects the appropriate subagent type based on task semantics, without needing hand-written routing rules in CLAUDE.md.
Possible implementation options
subagent_type: "auto"in the Agent tool — Claude infers the best agent type from task description + available agents- Declarative routing config — a structured section in
CLAUDE.mdor.claude/routing.jsonthat maps task patterns to agent types, evaluated automatically - Agent skill tags — agents declare their capabilities (
skills: ["code", "security"]), router matches based on task embedding similarity
Expected behavior
User: "Fix the bug in quota_manager.py"
→ Router detects: code change task
→ Automatically delegates to: impl-agent (Opus)
User: "Explain how the circuit breaker works"
→ Router detects: analysis/explanation task
→ Automatically delegates to: analyst-agent (Sonnet)
Why this matters
Multi-agent workflows are powerful but currently require significant orchestration boilerplate. Native routing would make it easier to build robust, specialized agent teams without maintaining complex routing logic in every project's CLAUDE.md.
Current workaround
Manual routing rules in CLAUDE.md + explicit Agent tool calls. Works, but fragile and project-specific.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗