| name | sgr_pool |
| description | SGR-enhanced pool of cheap specialized subagents based on sgr-agent-core. Uses Schema-Guided Reasoning to boost mini/Flash models to near-Sonnet quality. Trigger: "sgr", "sgr pool", "cheap agent", "mini agent", "sgr-researcher", "sgr-verifier", "sgr-summarizer".
|
sgr_pool — SGR-Enhanced Agent Pool
Purpose
A pool of specialized subagents that use sgr-agent-core as a Python library
for executing research and analytical subtasks with cheap LLMs (gpt-4.1-mini, gemini-2.5-flash, claude-haiku-4-5)
at quality approaching expensive models, thanks to Schema-Guided Reasoning.
Agent Types
| Agent | Task | Default model |
|---|
sgr-researcher | Search, information gathering, answering questions | gpt-4.1-mini |
sgr-verifier | Fact-checking, claim verification | gpt-4.1-mini |
sgr-summarizer | Summarization, compression | gemini-2.5-flash |
Usage
Python API
from skills.sgr_pool.executor import SGRPoolExecutor
executor = SGRPoolExecutor()
result = await executor.run("sgr-researcher", "What are the main approaches to knowledge graph construction?")
print(result.answer)
print(result.sources)
print(result.reasoning_steps)
result = await executor.run("sgr-verifier", "RAG always improves LLM accuracy", model="claude-haiku-4-5")
result = await executor.run("sgr-researcher", "What do we know about knowledge graph architecture?", use_kg=True)
Routing Rule
An SGR agent is selected by the orchestrator when:
- The task is classified as
research | verify | summarize
- Estimated complexity <
medium
- No file creation or code generation is required
For complex multi-step research → use AutoResearch (API-first).
For code generation → use CLI-first subagents.
Available Agent Tools
- WebSearchTool — Tavily search (built into sgr-agent-core)
- ExtractPageContentTool — page content extraction (built-in)
- KnowledgeGraphTool — BM25+keyword search across
02_distill/ (custom)
- WorkspaceReadTool — workspace file reading (custom)
Configuration
Agent configs: agents/*.yaml
Each yaml defines: model, tools, system prompt, execution limits.
Dependencies
sgr-agent-core>=0.7.0 (pip)TAVILY_API_KEY env var (optional, for web search)- Any OpenAI-compatible API endpoint
🇺🇦 Українською
Пул спеціалізованих субагентів на базі sgr-agent-core, що використовують Schema-Guided Reasoning для підсилення дешевих LLM (mini/Flash) до якості, наближеної до Sonnet. Три типи агентів: researcher, verifier, summarizer.
