Agent Tier Routing Guide
3-tier agent system: LITE → MID → MAX
Decision Flow
- LITE first for simple/cheap tasks
- MID for normal complexity
- MAX only when MID fails 3+ times
Per-role routing
Architecture
| Task |
Agent |
| Simple design sketch, brief review |
@architect-lite (LITE) |
| Full architecture, domain modeling, DB schema |
@architect (MID) |
| Unsolvable architecture problem |
@escalation (MAX) |
Code Analysis
| Task |
Agent |
| Quick file find, simple grep |
@context-manager-lite (LITE) |
| Deep codebase analysis, refactoring scope |
@context-manager (MID) |
| Beyond MID capability |
@escalation (MAX) |
Implementation
| Task |
Agent |
| DTO/Entity boilerplate, simple CRUD, basic test |
@worker-lite (LITE) |
| Business logic, API binding, complex test |
@worker (MID) |
| Beyond MID capability |
@escalation (MAX) |
Error Fixing
| Task |
Agent |
| Missing import, syntax fix, format lint |
@inspector-lite (LITE) |
| Multi-step compile error, runtime bug |
@inspector (MID) |
| 3+ failures on same bug |
@escalation (MAX) |
Cost Awareness
- LITE: deepseek-v4-flash / mimo-v2.5 — $0.14/$0.28 per 1M tokens
- MID: deepseek-v4-pro / mimo-v2.5-pro / qwen3.6-plus — $0.50~$1.74/$3.48
- MAX: glm-5.1 — $1.40/$4.40 (880 req/5hr limit, use sparingly)
Rule
Never use MAX for problems that LITE or MID can handle. Let the failed agent's prompt guide escalation.
pray4skylark