using-ai-frontier

AI Frontier

You have access to deep research intelligence via four data sources and four synthesis agents.

Critical: Distinguishing Foundational from Current SOTA

Citation count measures influence, not currency. A paper with 5000 citations from 2022 may be foundational but no longer SOTA. Always consider:

• Publication date — papers < 6 months old are current; > 2 years old need SOTA validation
• Supersession — has a newer method replaced this? (e.g., ReAct 2022 → superseded by ToT, LATS, ReWOO)
• Adoption signal — GitHub stars, HF upvotes, framework integrations indicate real-world use

When unsure whether a method is still SOTA, use perplexity-search --sota to get web-grounded current status.

Decision Tree

When you encounter these situations, use the appropriate tool:

Need to evaluate architecture against SOTA

Agent: architecture-evaluator Triggers: "Is this the right architecture?", "Evaluate my approach", "Are there better patterns for X?", "Review this against SOTA", "Is my architecture optimal?", "What patterns should I use for X?" Returns: SCOPE, CURRENT ARCHITECTURE (with ASCII diagram), SOTA ALTERNATIVES, GAP ANALYSIS, RECOMMENDATIONS, REFERENCES

Need to understand what the field knows

Agent: literature-reviewer Triggers: "What's the SOTA for X?", "Survey of approaches to X", "What does the research say about X?" Returns: DATA SOURCES, CONSENSUS, FRONTIER, OPEN QUESTIONS, KEY PAPERS, METHOD TAXONOMY, APPLICABILITY

Need to choose between approaches

Agent: method-analyst Triggers: "Should I use X or Y?", "Compare methods for Z", "Tradeoffs between approaches" Returns: COMPARISON MATRIX, RECOMMENDATION, IMPLEMENTATION NOTES

Need to implement a research method

Agent: implementation-guide Triggers: "How to implement X from paper Y?", "Code for method Z", "Architecture of approach X" Returns: CORE ALGORITHM, ARCHITECTURE, REFERENCE IMPLEMENTATIONS, ADAPTATION GUIDE

Quick lookups (skip agents, use skills directly)

Need	Skill	Example
Bleeding-edge preprints	arxiv-search	Latest papers on speculative decoding
Citations + TLDRs + impact	semantic-scholar-search	How influential is paper X?
AI summaries + trending	hf-papers-search	Trending papers, AI-extracted keywords
Current SOTA / what replaced X	perplexity-search --sota	What replaced ReAct for agent reasoning?
Last N days developments	perplexity-search --recent	New RAG frameworks this month
Web-grounded fact check	perplexity-search	Is paper X still considered SOTA?

When NOT to Use

• API docs, library usage, tutorials — use web search or documentation skills
• Common engineering knowledge — not a research question
• Debugging errors — use debugging tools

Data Sources

Free (no API key)

• arXiv — preprints with categories, dates, PDF links
• Semantic Scholar — citations, TLDRs, impact scores, venue info
• HF Papers — AI summaries, community upvotes, code repos, trending

Requires OPENROUTER_API_KEY (~$0.003/query)

• Perplexity — web-grounded SOTA discovery, recent developments, practitioner adoption

All retrieval skills return a unified JSON envelope. Agents return structured markdown sections.

Recommended Workflow for Research Questions

1. Academic search (arXiv + S2 + HF) → structured paper metadata
2. SOTA validation (Perplexity --sota) → what's current, what's superseded
3. Synthesis (agent) → merged analysis with both academic rigor and SOTA awareness

If Perplexity is unavailable (no API key), agents still work with academic data only — results will have citation-based ranking without web-grounded SOTA validation.