The Agent chooses how · Context controls where and what · tools run in parallel
Modern AI chat products are not a single LLM. An agent chooses tools, a context layer controls scope and policy, and retrieval or action tools run in parallel.
This OKPy special maps Agent vs Context, web search vs browser automation, nine RAG styles, and how Cursor, ChatGPT, Claude, and Gemini fit the same architecture.
❓
Question
The model does not know
📜
Long Context
Skip search · expensive
🧭
Agent
Judge · choose how
runs on LLM
Flash · Pro · GPT · Claude
How the Agent picks a tool
🧠
Memory
Dialogue · experience
⚙️
Context Layer
Gate · where · what
Meaning
What is being asked?
Clarify the request and narrow ambiguity
Permission
Where may it go?
Allow only visible data and systems
Policy
What is allowed?
Filter harm, secrets and personal data
Only passed scope ↓ to tools
PARALLEL TOOLS · choices at the same tier
Click a tool icon for details
RETRIEVAL
↕ same tier · parallel choices
ACTION · SENSE · SOLVE
Representative models · the Agent’s brain
Select a card to expand details
Model ≠ tool: Gemini Flash / GPT mini / Claude Haiku sit with the Agent. Web Search, RAG and Action sit in the tool tier below.
Agent & Context in detail
Select a card to expand details
🧭 Where tool-selection criteria live
The rules are not only “inside the model.” Developers provide descriptions, rules and routing.
In short: The Agent may choose the database, but Context can block it or narrow its scope.
9 RAG methodologies
Select a card to see details
9 other methodologies
Approaches that complement RAG or do not retrieve document chunks