From agent loops and tool use through MCP, ReAct, Reflexion, agent frameworks, multi-agent orchestration, computer-use agents, and Devin-style capstone. 17 lessons covering Anthropic's workflow patterns, CrewAI/AutoGen/LangGraph, SWE-bench-driven code agents, and production hardening.
What makes an AI agent different from a model? Augmented LLM, perception → reasoning → action loop.
How LLMs learn to use external tools through structured function calling. OpenAI tools, Anthropic tools.
Token-by-token generation of tool calls, constrained decoding, parallel tool execution, Pydantic validation.
Anthropic's open standard for tool/resource/prompt integration. The 'USB-C of AI' — write once, run with any host.
The most battle-tested agent pattern. Interleaving Thought → Action → Observation in a tight loop.
Short-term vs long-term memory, vector stores, MemGPT, episodic vs semantic vs procedural memory.
Verbal RL — agent attempts → critic evaluates → reflection note → retry. 10-30% quality lift on hard tasks.
Plan-and-Solve, Tree-of-Thought, LATS — when to escalate from ReAct to structured planning.
Anthropic's 5 patterns — prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer.
Compare LangChain/LangGraph, CrewAI, AutoGen, and Anthropic Agent SDK. Building blocks, role-based crews, conversation-driven patterns.
CrewAI / AutoGen / LangGraph. Supervisor and swarm patterns, framework comparison, production multi-agent orchestration.
Claude Computer Use API, OpenAI Operator, OSWorld benchmark. The most general-purpose agent interface.
Cursor, Devin, Claude Code, Copilot Workspace. SWE-bench architecture, repo-level reasoning.
SWE-bench Verified, OSWorld, GAIA, AgentBench. Trajectory vs outcome evaluation, 2026 reward-hacking caveats.
Constitutional AI, oversight, sandboxing, prompt injection defense, human-in-the-loop, the Swiss cheese model.
Langfuse / AgentOps tracing, OpenTelemetry, distributed tracing, cost monitoring, retries, fallbacks, the 12-point production checklist.
Capstone — Devin/Manus-style end-to-end agent. MCP + ReAct + memory + Reflexion + safety + observability + eval, integrated.
12 interactive labs — hands-on exercises for this track
An AI agent observes, thinks, acts, and observes again — watch the reasoning loop in real time
Watch an LLM call external tools — search the web, run code, check weather — and weave results into its response
Watch the Thought → Action → Observation loop unfold — the same LLM chains tool calls to solve multi-hop questions it couldn't answer in one shot.
Watch a team of specialized agents (Planner, Researcher, Writer, Coder, Critic) coordinate on a task. See how a critic gate rejects weak subtasks and sends them back for revision.
Three memory types in one view — short-term conversation buffer, long-term vector store with top-k retrieval, and task-scoped scratchpad. See how each solves a different problem.
Dissect the LLM function-calling pipeline — structured JSON, schema validation, dispatch, and result round-trip. See happy path, malformed JSON retry, unknown-tool refusal.
Watch an agent decompose a complex goal into a tree of verifiable leaves, then re-plan gracefully when a constraint fails.
Send adversarial and benign inputs through a 5-stage guardrails pipeline — prompt injection, PII, toxicity, hallucinations. Toggle to reveal what leaks without defenses.
Full 8-step JSON-RPC flow: initialize → capabilities → tools/list → tools/call. One client, 3 servers (GitHub, Filesystem, Postgres).
Side-by-side non-streaming vs streaming. Watch TTFT drop from 5000ms to 200ms — streaming FEELS 10× faster.
3 composite tasks (CSV→chart, research-summarize, support triage). See small composable tools beat one monolithic tool.
4 agents × 4 benchmarks (HumanEval, GAIA, SWE-bench, MMLU). See no single winner — cost-per-score and failure examples per benchmark.
425 questions across 17 modules — check how well you understood this track.