Area 03 / 08 · AGT
AI Agents
Phase 3 · Weeks 11–14Models that use tools in a loop until the job is done.
An agent is an LLM in a loop with tools: it decides what to do, does it, reads the result and decides again. The base concepts cover tool calling, the loop, planning, memory and MCP. The production half covers multi-agent patterns, frameworks, state, human approval, context management and stopping safely. Read Security 01–04 before you give an agent real tools.
0 of 20 used in code
01–10 · The base
Learn these first.
- 01Tool CallingThe model returns a function name + arguments. Your code runs it and sends the result back.
- 02The Agent LoopThink, call a tool, read the result, repeat until the task is done.
- 03ReAct PatternReasoning steps and actions take turns. The classic agent prompt pattern.
- 04PlanningBreak the goal into steps before acting, and update the plan as results come in.
- 05ReflectionThe agent reviews its own output and tries again to fix mistakes.
- 06Short-term MemoryWhat the agent holds in the current context: the conversation and recent tool results.
- 07Long-term MemoryFacts saved outside the context (files, database, vector store) and recalled in later sessions.
- 08MCP ServersModel Context Protocol servers offer tools, resources and prompts in one standard format.
- 09MCP ClientsThe app side (IDE, chat app, your agent) that connects to MCP servers and uses their tools.
- 10Agent SkillsFolders of instructions and scripts that the agent loads only when a task needs them.
11–20 · Production level
Come back after you have shipped something.
- 11Multi-agent HandoffsOne agent passes the task, with its context, to a more specialised agent.
- 12Orchestrator-Worker PatternA lead agent splits the work, workers do the parts in parallel, the lead merges the results.
- 13LangGraphA framework for building agents as graphs of steps with shared state, branches and loops.
- 14OpenAI Agents SDKA light framework for agents with tools, handoffs, guardrails and tracing built in.
- 15State and CheckpointsSave the agent’s state at every step so a run can pause, resume or be replayed.
- 16Human in the LoopThe agent stops and waits for a person to approve or edit before a risky step.
- 17Context EngineeringChoosing exactly what goes into the context at each step: instructions, tools, data, history.
- 18Context CompactionSummarise or drop old history when the context gets long, and keep what still matters.
- 19Stop ConditionsStep limits, budgets, timeouts and ‘done’ checks, so the agent can’t loop forever.
- 20A2A ProtocolAgent2Agent: an open protocol that lets agents from different teams find and talk to each other.
BUILD THIS
A research agent with web search and page-fetch tools, a written plan, a step limit, notes saved as memory, and one MCP server you wrote yourself.
Project guide →