Schedule at a glance

Week 1. Foundations: workflows, AI workflows, agents
Mon Oct 12 – 09:00–12:00 New York

Week 2. Multi-agent architectures & collaboration
Mon Oct 19 – 09:00–12:00 New York

Week 3. Interpretability, safety & decision tracing
Mon Oct 26 – 09:00–12:00 New York

Week 4. Tool-using & coding agents
Mon Nov 2 – 09:00–12:00 New York 

Week 5. Knowledge-integrated agents & RAG 2.0
Mon Nov 9 – 09:00–12:00 New York 

Week 6. LLM agents + RL layers in financial simulation
Mon Nov 16 – 09:00–12:00 New York

Week 7. Evaluation, governance & controlled pilots
Mon Nov 23 – 09:00–12:00 New York

Each session is followed by an optional 60-minute office hour.

The Curriculum: 09:00–12:00 New York Time

Mon Oct 12 · 3 hours – Miquel Noguer i Alonso
Week 1. Foundations: workflows, AI workflows, agents

Deciding what should be an agent — and what should stay a script. You ship: a running agent scaffold and a written rule for when not to use one.

Topics: The three tiers — deterministic workflow, AI-augmented workflow, autonomous agent — and the cost, latency and risk
each one buys you. Raw kernel versus agent kernel, and why the harness is the object you can specify, pin and test. Agent
anatomy: planning, memory, tool calls, stop conditions. Token and cost budgeting before a line of orchestration is written.
Choosing between Hermes, OpenAI Responses SDK, AutoGen, CrewAI, LlamaIndex and LangChain on grounds other than
fashion.

Lab: Stand up the course repository and run a single-agent research assistant on the Hermes harness that pulls market data
and drafts a one-page note — then instrument it with cost, latency and failure counters and watch what a careless loop
costs.

Deliverable A working agent scaffold, a cost dashboard, and a workflow-versus-agent decision rule written for your own desk.

---

Mon Oct 19 · 3 hours – Miquel Noguer i Alonso
Week 2. Multi-agent architectures & collaboration

When a crew beats a single agent — and when it just multiplies the bill. You ship: a three-agent investment committee that produces a defensible memo.

Topics: Crew roles, dynamic spawning, delegation and structured debate. Communication schemas, reflection and voting. Typed
hand-offs so agents cannot silently corrupt each other’s state. The failure modes of crews: consensus collapse, runaway
delegation, and cost blow-up.

Lab: Build a three-agent investment committee analyst, sceptic, chair that parses SEC filings and issues a decision memo with the dissent recorded rather than averaged away.

Deliverable: An investment-committee crew, its memo template, and a comparison against the single-agent baseline on cost and
quality.

---

Mon Oct 26 · 3 hours – David Pacheco Aznar
Week 3. Interpretability, safety & decision tracing

Making an agent’s decisions legible to someone who has to review them. You ship: an advisory agent whose every decision leaves a readable trail

Topics: Structured decision records capture tool calls, source traces, policy-check results and exceptions without exposing hidden reasoning. Automatic redaction keeps client data and material non-public information out of the record. Pausable
workflows, explicit human hand-back points and reviewer-facing audit trails.

Lab: Instrument an advice agent with structured decision records and automatic redaction; verify its tool and source traces,
policy checks and exception log; then test that redaction and human hand-back survive an adversarial prompt.

Deliverable A traced, redacted advisory agent and a structured decision-record format for your institution.

---

Mon Nov 2 · 3 hours – David Pacheco Aznar
Week 4. Tool-using & coding agents

Agents that write, run and repair code without escaping the sandbox. You ship: a self-repairing code agent and one MCP server exposing a desk tool

Topics: Coder–Critic–Executor loops (Reflexion, Self-Refine) and where each one stalls. Safe sandboxing, resource limits, and unittest- driven repair as the stopping criterion. Stateless-parser boundaries: empty toolset whitelist, no persistent memory,
strict JSON-only output at every proposal hand-off. MCP for exposing internal data and pricing tools to an agent without
handing over the keys.

Lab: A repair-bot iteratively fixes a faulty VaR calculator against a failing test suite; you then wrap a redacted copy of one of
your own pricing or reporting functions as an MCP server and let the agent call it under permission control.

Deliverable: A working repair loop plus an MCP server that exposes a real desk tool safely.

---

Mon Nov 9 · 3 hours – Miquel Noguer i Alonso
Week 5. Knowledge-integrated agents & RAG 2.0

Grounding agents in filings and internal documents without inventing citations. You ship: a streaming covenant monitor with citations that resolve to the page

Topics: Corrective-RAG (CRAG), introspective agents and structured retrieval over filings. Stateful agents with Letta. Streaming pipelines, vector-database operations and compliance logging. Chunking and citation discipline that preserves the audit
trail — and treating every retrieved document as untrusted input.

Lab: Build a streaming RAG agent that ingests live 10-Q filings and flags covenant breaches, then plant an adversarial instruction inside a filing and confirm the agent reports it rather than obeying it.

Deliverable: A covenant-monitor agent, its corpus configuration, and a document-poisoning test result.

---

Mon Nov 16 · 3 hours – David Pacheco Aznar
week 6. LLM agents + RL layers in financial simulation

Where reinforcement learning genuinely improves an agent and where it does not. You ship: reproduced crew-versus-PPO results, one substitution of your own, and a limitations note

Topics: Run a supplied frozen crew-versus-PPO benchmark with fixed data, costs, risk limits and evaluation window. Compare
Sharpe, CVaR, compliance-violation rate, latency and cost, separating orchestration effects from policy effects. The exercise is a controlled simulation, not evidence of production readiness.

Lab: Run and reproduce the supplied crew-versus-PPO benchmark in FinRL-Meta. Then substitute one component of your own — your cost model, your risk limits, or a signal your desk actually uses — rerun it under the same frozen window, and report
where the crew wins, where PPO wins and what remains outside the benchmark.

Deliverable: Reproduced benchmark results, one controlled substitution using your own cost model or signal, and a reviewer-facing limitations note.

---

Mon Nov 23 · 3 hours – Miquel Noguer i Alonso
Week 7. Evaluation, governance & controlled pilots

Deciding whether the system has enough evidence for a controlled pilot. You ship: your evaluated capstone and a controlled-pilot evidence pack

Topics: Harness validation separates replicate dispersion from instruction sensitivity; lock the design before viewing outcomes. Benchmarks include AgentBench and SWE-agent regression; risk metrics include CVaR, stress scenarios and failure rate. Controlled-pilot controls cover guardrails, kill-switches, budgets, monitoring, hand-back and EU AI Act documentation.

Lab: Run the capstone through the evaluation harness, red-team it, and build the monitoring dashboard and audit procedures required for a controlled pilot.

Deliverable: An evaluated capstone plus the controlled-pilot evidence pack: evaluation results, red-team report, monitoring plan and a reviewer-facing limitations note.