Master AI agents in finance end to end, and make them production-grade

One research agent built three ways: the raw loop, the SDK, the graph. Around it, the full design-pattern map and the lens the whole course applies: in production, failure is the default case, cost is a budget, and every run must be reconstructable.

Key concepts: the agent loop, one agent three ways, the design-pattern taxonomy (capability, reasoning, quality, orchestration), what production changes.

1 lab: the same research agent three ways, each variant with a production upgrade the tutorials skip.

How agents decide what to do before doing it, and the workflow patterns behind reliable pipelines: when a fixed, inspectable flow beats an autonomous agent, and how to run specialists in parallel.

Key concepts: planning styles and their trade-offs, prompt chaining with gates, parallelization, routing, orchestrator-workers, evaluator-optimizer, workflows vs agents in finance.

2 labs: a planning agent that survives a compliance surprise mid-run, and a parallel impact assessment across four types of finance actors.

The signature chapter of this level, and a discipline no other level teaches: the harness is everything the loop must own that the model doesn't. Runaway loops, silent failures, and cost blowouts each get a detection signal and a guard.

Key concepts: turn and token budgets, stop conditions, tool dispatch, retries and timeouts, cost circuit breakers, checkpoint and resume, human interrupts, the loop failure taxonomy.

1 lab: familiar agent loops re-read through production robustness: bounded retries, budget guards, and disagreement as a live escalation signal.

The orchestration map, then the production version: what state agents share, where a human signs off, and what a team needs to survive real traffic. Ends with a framework-landscape survey and a decision checklist.

Key concepts: supervisor/workers, orchestrator-worker, handoffs, agents-as-tools, shared state, context isolation, hardening a team, choosing a framework.

1 lab: a hardened supervisor team you keep all course: instrumented in week 5, evaluated in week 5, served in week 6.

The complete memory picture: what each kind of memory enables, how an agent learns from its own past sessions, and the production question underneath: what should an agent be allowed to remember, and who reviews it?

Key concepts: working vs long-term memory (episodic, semantic, procedural), learning from past sessions, compaction policies, pinned constraints, memory as a governance surface.

2 labs: an advisor that distills sessions into retrievable experience, and a compression lab measuring what each policy forgets and what it costs.

The full reasoning landscape, deep this time: how agents check their own work, and the search-based patterns that explore many candidate answers before committing, each with an honest cost model.

Key concepts: ReAct, self-refinement and reflection, self-consistency, tree-of-thoughts, tree search with backpropagation, verification patterns, LLM-as-a-judge with rubrics.

3 labs: a self-refining investment thesis writer, a tree-of-thoughts hedging decision, and a tree-search agent rebalancing a portfolio under constraints.

Keeping an agent inside its mandate: guardrails as first-class components, the subtle race condition where a tool fires before the guardrail verdict lands, and red-teaming your own system before someone else does.

Key concepts: input and output guardrails, tripwires, the guardrail-vs-tool race condition, compliance-safe behavior, red-team batteries as living test assets.

2 labs: harden a fundamentals agent against out-of-scope queries and the race condition, then attack an advisory agent with a red-team battery.

Agents that touch real workflows need defense and a paper trail: resisting injected instructions, tiering actions by how reversible they are, and being able to reconstruct for a reviewer exactly why the agent did what it did.

Key concepts: prompt-injection defense, action tiers, least-privilege tools, the trace as an audit file, tamper-evident plans, reconstructing an agent's history from its records.

2 labs: a hardened client-intake agent, and an audit lab that reconstructs an agent's decisions end to end.

From reading traces to running monitoring: instrument a whole agent team persistently, attach cost to every span, and turn a pile of runs into a dashboard someone can act on.

Key concepts: traces and spans for teams, persistent instrumentation, cost accounting per agent and per run, latency budgets, monitoring vs looking.

1 lab: instrument the week-2 team end to end and build its cost dashboard.

The chapter that separates a demo from a system: how do you know the agent is good, and how do you know it's still good after you change it? Judges where judgment is needed, code where code is better, and thresholds you re-check after every change.

Key concepts: offline evals over traces, synthetic datasets, LLM-as-a-judge vs deterministic code evaluators, tool-selection evals, trajectory convergence.

4 labs: judged QA at scale, tool-call evaluation, trajectory convergence, and code-execution checks.

First, what agentic RAG actually is: retrieval that routes, decomposes, grades itself, and refuses when the evidence isn't there. Then the part most courses skip: measuring it, with the metrics that tell you whether the retriever or the generator is failing.

Key concepts: naive-RAG failure modes, routing and decomposition, retrieval grading, cite-or-refuse, the RAG evaluation triad (faithfulness, answer relevance, context relevancy), hallucination metrics, ground-truth metrics.

2 labs: a metrics tour where a deliberately wrong answer makes the scores fail visibly, then build a routed retrieval pipeline over a real filing and evaluate it end to end.

Week 7 assembles everything into three production systems: an Autonomous Rates Analyst (Iron Reflex), a Regulated Robo-Advisor Platform (By the Book), and a Global Macro Monitoring Desk (Watchdog). Then you design and build your own capstone: at least two cooperating agents, with guardrails, tracing, and an eval. Full details below.