LearnAISign in
Phase 10 · 12 lessons · 39 h · about 4 weeks

🤖 Agentic AI

Models that act: tools, memory, planning, protocols, evaluation and safety

Lessons

10.1What is an agent? Loops, autonomy and workflows vs agentsThe perceive–reason–act loop, and when not to build one2 h 30 min10.2Tool use and function callingSchemas, the tool-result loop, errors, code execution and computer use⚡ demo · 3 h10.3Reasoning and planning for agentsReAct, plan-and-execute, search over thoughts, reflection, reasoning models⚡ demo · 3 h10.4Memory and context engineeringThe context window as working memory, compaction, notes, retrieval and long-term stores3 h10.5Agentic workflow patternsChaining, routing, parallelization, orchestrator–workers, evaluator–optimizer2 h 30 min10.6Multi-agent systemsTopologies, communication, the share-context debate, failure taxonomy and frameworks2 h 45 min10.7MCP and interoperability (plus Agent Skills and A2A)Hosts, clients, servers; tools, resources, prompts; code execution; skills; agent-to-agent2 h 45 min10.8Coding agents and computer useWhy code came first, agent–computer interfaces, harnesses, GUIs as environments2 h 45 min10.9Evaluating agentsTrajectories, graders, pass@k vs pass^k, compounding error, benchmarks and harnesses⚡ demo · 2 h 45 min10.10Agent safety and securityPrompt injection, the lethal trifecta, least privilege, sandboxing and human-in-the-loop2 h 45 min10.11Building production agentsState machines, durable execution, harness design, observability, cost control and release3 h10.12Project: build a research agent with tools, an MCP server and an eval harnessA capstone that is evaluated, instrumented, budgeted and safe by design8 h

About this phase

An agent is a language model placed inside a loop: it reads the state of the world, reasons, acts through tools, observes the result and repeats until the job is done. This phase turns everything you learned about LLMs (8.1 The LLM lifecycle and model landscape–8.7 Evaluating LLMs) and retrieval (9.1 Why RAG, and the canonical pipeline–9.11 Project: a cited RAG assistant over your own documents, with evals) into systems that do things: call APIs, write and run code, browse, remember across sessions and coordinate with other agents. You will build the loop by hand, study the planning and memory techniques that make it reliable, learn the protocols that connect agents to tools and to each other (MCP, Agent Skills, A2A), and treat evaluation and security as first-class engineering — because an agent that is right 95% of the time per step is wrong most of the time after twenty steps. The phase ends with a research agent you design, instrument and evaluate yourself.

By the end you can

  • Implement a tool-using agent loop from scratch against a frontier model API and explain every message in the transcript
  • Choose between workflows, single agents and multi-agent systems using cost, reliability and verifiability arguments — and defend the choice
  • Design memory and context strategies (compaction, notes, just-in-time retrieval, long-term stores) for long-running tasks
  • Build and connect an MCP server, package procedural knowledge as Skills, and reason about the security of tools and protocols
  • Evaluate agents with trajectory-level harnesses, pass@k vs pass^k, and compounding-error math, and harden them against prompt injection and the lethal trifecta
  • Ship a production-grade agent with explicit state, durable execution, observability, budgets and human-in-the-loop controls