Skip to the content.

Abstract

This note provides a practical “capability map” for designing agentic systems: for each target capability, it lists (1) the typical limitation in current GenAI systems, (2) an engineering substitute, (3) common implementation techniques, and (4) the residual gap that remains.

The goal is not neuroscience completeness. The goal is fast, explicit tradeoff reasoning when you design memory, retrieval, planning, and safety layers around an LLM.

Data (CSV)

How to use this map

1) Scope your agent

2) Choose the substitute layer

3) Treat “Residual gaps” as design risk Residual gaps are where production systems typically fail (consistency, provenance, safety, interpretability). Track them as risks, not as footnotes.

What this table is (and is not)

Suggested next read in this repo