Syllabus – 21 Lecture Hours

This intensive, hands-on course explores cutting-edge Large-Language-Model (LLM) Agents-and their hybridisation with reinforcement learning-in the financial domain. Students design, orchestrate, interpret and govern agentic systems for algorithmic trading, fraud detection, risk management and personalised advisory. Every session features sub- stantial coding labs using production frameworks (OpenAI Responses SDK, AutoGen, CrewAI, LlamaIndex, LangChain) alongside emerging RL– tool-use libraries (ARTIST, ReCall, verl).

SESSION 1 Foundations – Workflows – AI Workflows – Agents (3 h)

• Working knowledge of Python programming.
• Introductory understanding of financial markets and instruments.
• Basic familiarity with machine-learning concepts; prior RL is helpful but not required.

SESSION 2 Multi-Agent Architectures & Collaboration Patterns (3 h)

• Crew roles, dynamic spawning, delegation and debate.
• Communication schemas, reflection, voting.
• Lab: three-agent investment committee parses SEC filings and issues a decision memo.

SESSION 3 Interpretability, Safety & Thought Tracing (3 h)

• Anthropic circuit tracing; scratch-pad logging & sanitisation.
• Governance: audit trails, pausable workflows.
• Lab: instrument advice agent with circuit traces and auto-redaction.

SESSION 4 Tool-Using & Coding Agents (3 h)

• Coder–Critic–Executor loops (Reflexion, Self-Refine).
  Safe sandboxing and unit-test-driven repair.
• MCP and advanced tools
• Lab: repair-bot iteratively fixes a faulty VaR calculator.

SESSION 5 Knowledge-Integrated Agents & RAG 2.0 (3 h)

Corrective-RAG (CRAG), Introspective Agents, structured retrieval.
Stateful agents with Letta
Streaming pipelines, vector-DB ops, compliance logging.
Lab: streaming RAG agent ingests live 10-Q filings and flags covenant breaches.

SESSION 6 LLM Agents + RL Layers in Production Finance (3 h)

• RL Adds Value: RLHF/RLAIF and policy-level RL (ARTIST) for tool-selection and compliance.
• Turn-level credit assignment; ReCall & Agent-R1 for end-to-end tool-calling RL.
• Hybrid market simulators: TradingAgents crews, FinRL-Meta;
• LLM-only vs PPO trader comparison.
• Evaluation: AgentBench finance tasks, CRAG-based factual rewards, Self-Taught Evaluator.
• Lab: four-agent prop-desk crew (planner, analyst, coder, risk officer) deployed in FinRL-Meta equities env-targets: Sharpe, CVaR, compliance.

SESSION 7 Evaluation, Governance & Deployment (3 h)

• Benchmarks: AgentBench, SWE-agent regression.
• Risk metrics: CVaR, stress scenarios, fail-rate.
• Deployment pipelines: blue–green, guard-rails, kill-switches, EU AI Act mapping. Capstone: monitoring dashboards and audit proced