Cheat sheetAIS-01

Agentic AI Foundations

AI Specialization / Agentic AI Foundations

An agent is an LLM that runs a loop over tools with memory, choosing its own next action until a goal is reached. Know the loop, and know when NOT to build one.

01
Four ingredientsModel (reasons) + Tools (act) + Memory (carry context) + Loop (feed results back and repeat until a stop condition).
02
ReActInterleave Thought -> Action (tool call) -> Observation (result), then reason again. Adaptive but can wander without limits.
03
Plan-and-executeWrite a full plan first, then run the steps, re-planning only on failure. Cheaper and more predictable than ReAct.
04
Agent vs workflowAgent = model decides control flow at runtime. Workflow = you hard-code the steps. Prefer the simplest design that works.
⚑

For your next LLM feature, first ask: does the path branch unpredictably? If no, build a workflow. If yes, build an agent and give it explicit tools, a memory strategy, and a hard stop condition (max steps + budget + success check).

Agentic (needs a loop)"Investigate why the nightly job failed and fix it" — steps are unknown until the agent reads logs, so let the model drive.
Not agentic (use a workflow)"Summarize this ticket, then classify its priority" — two fixed steps; a hard-coded pipeline is cheaper and safer.
loop: model -> pick tool -> run tool -> observe -> repeat
stop when: goal met OR max_steps hit OR budget exceeded
ReAct step = Thought + Action + Observation
agentsreactplan-executeautonomyworkflow-vs-agent
review in 6d