Financial analyst agent

What you'll build

• Five market-data tools for ticker lookup, price performance, fundamentals, news and market movers, none of which needs a market-data key.
• Compact tool output that summarises a year of daily prices in code before the model sees it.
• A scope written into the system prompt as two flows and a polite refusal, checked on one request of each kind.
• Follow-up questions on one thread, with the numbers in the answers checked against the provider.

Lesson 2 of T8.3's agents answered from web pages, where a claim can only be traced to a URL. This agent answers from structured market data, where every number has a source to compare against. The machinery is unchanged: a create_agent agent from T7.2, T7.1-style tools and an T4.4 checkpointer for follow-ups.

What changes is the discipline around numbers. The model should read figures and never compute them, a missing value must say it is missing, and the agent must decline questions outside its job. On Larkspur the same shape would fit a billing analyst over invoice data. The notebook is an engineering exercise, not investment advice.

The agent sits between an investor and five tools, each backed by a free data provider:

Figure 1: The analyst and its tools. OpenBB routes each market-data call to a provider, and news comes from a web search.

Table 1: The project's pieces.

The price tool is where the arithmetic moves out of the model. Here is what each design hands over per call:

Table 2: One year of prices, two ways. Returns use 21 trading days per month.

The notebook checks the scope with one request per path, and passes each on what the agent called, not on its wording:

Table 3: Three planted requests and their pass conditions.

The misspelled request exercises the recovery rule in the prompt. The values change daily, and the trace shows only the call order:

Figure 2: A failed lookup becomes a hint the agent acts on, then the ticker feeds the next call.

The follow-up conversation then asks What is the P/E ratio of Nvidia?, And Intel? and Which of the two looks better on valuation? on one thread. The check fetches the same metrics from the provider directly. Nvidia's answer must contain its trailing P/E, rounded to one or two decimals. If the provider has no trailing P/E for Intel (the notebook was written while Intel's trailing earnings were negative), the answer must say so or quote the forward P/E, and any other P/E figure counts as invented.

Table 4: Six choices and what each one protects against.

• Live data. Every run sees different prices. The checks test agreement with the provider at run time, never fixed values.
• Scope lives in the prompt. Nothing in code blocks an off-topic tool call, and one request per path is a sample. A determined user can talk the agent around.
• A loose number match. The check accepts any number within 0.051 of the provider's value, so an unrelated figure that happens to be close also passes.
• Advice. The prompt allows recommendations with a do-your-own-research caveat. A production version would need compliance review.

The notebook needs OpenAI and Tavily keys and the course's finance dependency group, which cannot share an environment with the protocol groups.

The code for this lesson is in these notebooks. Colab runs one in your browser: install the packages its first cell imports and set the API keys it reads. To run every notebook as written, clone the repository and use its uv environment.

• Do the arithmetic in the tool and hand the model a short table, so it reads figures instead of computing them from hundreds of rows.
• Make missing values explicit in tool output, and treat any figure the provider did not return as invented.
• Test a scoped agent on one request per path, judging each by the tools it called, and check its numbers against a direct fetch from the same provider.