| name | routr-agents |
| description | Design multi-agent systems, agent loops, and eval harnesses. Use when: building agents, agent operating loop, research/self-correction loop, harness, memory, tools, context engineering, eval loop. Not for: a single AI SDK chat/RAG feature (→ routr-ai). |
routr-agents
Goal: principled agent architecture — not ad-hoc prompt stacking.
When to activate
- Building agents, multi-agent systems, agent operating/research loops
- Eval harness, context engineering, memory, tool/MCP design
Do not activate
- Single AI SDK chat/RAG feature, no architecture question →
routr-ai
- Agent UI (chat window, controls) →
routr-frontend
Iron law
Prototype the harness on one path before scaling prompts — never scale a pattern that hasn't been proven once.
0. Bootstrap
See routr-catalog/references/skill-registry.md. Load by sub-task only — never read the whole cluster.
Context engineering
| Sub-task | Skill |
|---|
| Concepts | context-fundamentals |
| Degrading mid-session | context-degradation |
| Long handoff | context-compression |
| Token budget | context-optimization |
| File-based context | filesystem-context |
| Memory | memory-systems |
| Tools / MCP | tool-design |
Loop engineering (the operating loop — growing fast, pick per symptom)
| Sub-task | Skill |
|---|
| Design the agent operating / research loop | harness-engineering |
| Robust research loop (validation, error recovery) | comprehensive-research-agent |
| Debug a misbehaving loop (drift, tool confusion, repeat failures) | reasoning-trace-optimizer |
| Eval loop / quality gate | evaluation, advanced-evaluation |
| Multiple agents / handoffs | multi-agent-patterns |
| Subagent delegation loop | subagent-driven-development |
| Background / hosted / self-spawning loop | hosted-agents |
| Cross-agent memory briefing | latent-briefing |
Runtime: lean-ctx when implementing.
Budget gate: read one row per active symptom. Do not preload the context + loop clusters together.
Skip: frontend-design, caveman unless terse docs requested.
1. Clarify
Single vs multi-agent; human-in-loop; context budget; eval criteria.
2. Design
Pattern → tools → memory → harness.
3. Implement
Prototype harness before scaling prompts.
Output format
## Agent design report
**Pattern:** single / multi-agent, loop shape
**Context, memory, tools:** …
**Prototype status:** proven on one path / not yet
Handoff
- Wrong file reads →
routr-debug + symdex
- Agent UI →
routr-frontend
References
