Research-Before-Plan
Benchmarked against: Anthropic โ Adaptive thinking Source: Agent Intelligence Protocol ยง6 + Company Constitution ยง9
Never plan with stale knowledge. This protocol ensures every technical decision is based on verified, current information โ not training data that may be months old.
The workflowโ
Knowledge classificationโ
| Category | Examples | Action |
|---|---|---|
| Timeless | Design patterns, algorithms, SOLID principles, SQL syntax | Use directly from training |
| Perishable | Library versions, API signatures, framework conventions, cloud service configs | Verify before using |
The search cascadeโ
- Search UB first (
search_brain) โ Free, instant, team knowledge - UB empty or stale โ Delegate to external search
- Groq (
search_webengine=groq) โ Free, fast - Gemini (
search_webengine=gemini) โ Authoritative, has citation links
- Groq (
- Ingest results (
ingest_fragment) โ Now it's a company asset - Then plan โ Based on verified knowledge
Why this mattersโ
Real incidents from skipping Research-Before-Plan:
| Incident | Root cause | Impact |
|---|---|---|
| "LangGraph 2.0 Released" intel | Groq hallucinated version number | Polluted UB with false information |
| 3 WOs rejected same day | No Pre-Flight Check on perishable knowledge | Wasted agent and Captain time |
| Deprecated API in production code | Wrote code from training data memory | Runtime failures |
Integrationโ
- Pre-Flight Check: Knowledge Freshness dimension (score 2 = must verify)
- Tech Freshness: Danger Zone library list identifies high-risk packages
- Cost Awareness: Search delegation follows the engine cost hierarchy