Agentic AI Cheat Sheet

Interleave explicit step-by-step reasoning thoughts with action tool calls and observation inputs.

First phase generates structured DAG plan breakdown; second phase executes sub-tasks sequentially.

Evaluator LLM checks output against acceptance criteria, returning feedback to generator until pass.

Pause agent execution loop on high-risk tool calls (e.g. write file, run terminal command) awaiting user approval.

Declare tool name, description, and strict parameter properties/types in system schema payload.

Standard protocol method invocation: `tools/call` with parameters over stdio IPC or SSE transport.

Format tool outputs clearly as string or JSON observations injected directly back into chat context.

Central supervisor agent inspects goal, routes tasks to domain subagents, and synthesizes output.

Tree hierarchy where department sub-supervisors manage worker agents for deep sub-tasks.

Linear chain where Output of Agent A becomes Input context for Agent B (e.g. Specs -> Code -> Test).

Hard cutoff (e.g. max_steps = 15) preventing runaway infinite tool loops or budget exhaustion.

Restrict write/execute tool access to specific workspace directory paths, blocking root/system access.

Prune old observation logs or summarize earlier loop turns when approaching model context window limits.