Memory, tools, and prompts
Memory: working vs archival
Not all memories need to be fast. Your name sits in working memory — accessed every turn. Your birthday goes to archival memory — needed once a year. Working memory must be highly optimised because slow retrieval creates latency on every single interaction.
Tools: APIs and beyond
An agent's tools are the external systems it can call: flight search APIs, hotel booking systems, payment processors, weather services. LLMs are surprisingly good at reading API documentation and figuring out how to call an endpoint if you give them the spec.
Degrees of autonomy
Least autonomous: the steps are hard-coded (always do A then B then C). More autonomous: the tools are defined but the agent decides the order. Most autonomous: the agent decides the steps AND can create new tools — writing code, calling any API, performing calculations.
The travel agent example
A travel booking agent combines all three: memory (user prefers direct flights, budget hotels), tools (flight API, hotel API, payment API), and prompts (plan the trip, validate with the user, update preferences). After each interaction, it updates memory for next time.
If a travel agent AI forgot your preference for direct flights every session, how would that affect your experience?