32 lines
1.4 KiB
Markdown
32 lines
1.4 KiB
Markdown
# Agent Runtime Policy (NODE1)
|
|
|
|
## Purpose
|
|
Single policy for runtime model selection and orchestration behavior.
|
|
|
|
## Agent Classes
|
|
- `top_level`: Telegram-facing orchestrators with optional CrewAI teams.
|
|
- `internal`: infrastructure/service agents (not Telegram-facing by default).
|
|
|
|
## Model Policy
|
|
- Top-level agents: `cloud_deepseek` primary + `cloud_mistral` fallback.
|
|
- Exception: `sofiia` uses `cloud_grok` primary + `cloud_deepseek` fallback.
|
|
- Internal agents: local-first (`ollama` profiles).
|
|
- `monitor`: `qwen2_5_3b_service`.
|
|
- `devtools`: local by default; cloud override only for explicitly heavy tasks.
|
|
- `comfy`: no chat LLM profile (tool/service execution path).
|
|
|
|
## Orchestration Policy (CrewAI)
|
|
- Top-level agents are direct-LLM first for simple requests.
|
|
- CrewAI is on-demand for complex/detailed requests (`force_detailed` / `requires_complex_reasoning`).
|
|
- Fast path must cap active subagents to 2-3 roles.
|
|
- Final user response is always formed by the top-level agent.
|
|
|
|
## Routing Constraints
|
|
- Do not use code-level hard overrides for cloud/local policy if it can be expressed in `router-config.yml`.
|
|
- Prefer deterministic routing rules per agent over generic fallbacks.
|
|
|
|
## Source of Truth
|
|
- Canonical intent: `config/agent_registry.yml`.
|
|
- Runtime implementation: `services/router/router-config.yml`.
|
|
- Any policy change must be reflected in both until router config generation is fully automated.
|