What does an investing agent actually do?

Define the job before judging the answer

Begin with a task that has a visible stopping point: compare two dated disclosures, identify missing evidence, or explain how a written rule would handle a historical case. A general instruction to find good investments leaves the information boundary, risk policy, and success criteria unresolved.

Consider a hypothetical research assistant asked to review whether an invented company meets a disclosure checklist. It retrieves the permitted documents, maps each requirement to a passage, marks two items unknown, and returns a draft. The useful outcome is a record another person can challenge. A confident buy label would add a conclusion that the task never established.

Separate three kinds of output

A summary describes available material. A simulation applies chosen rules and assumptions to a bounded input. An execution instruction asks an external system to change something. Treating these as separate jobs makes the permission and evidence requirements easier to inspect.

• Research: What source supports this statement, and when was it available?
• Simulation: Which rule version, data coverage, and fill assumptions produced this result?
• Execution: Which person or explicit policy can authorize a real action, and how is that authority enforced?

Use a small evaluation record

For one hypothetical task, save the exact question, permitted sources, cutoff time, agent version, observed output, and unresolved errors. Ask a second reader to reproduce the source check without seeing the first conclusion. Disagreement is valuable evidence about ambiguous definitions or missing context.

Include a case where the required document is missing. A useful assistant should preserve that gap rather than inventing a passage. Also include a source that contradicts the preferred thesis. These checks evaluate behavior on specific cases; they do not demonstrate general investing competence.

Read capability claims narrowly

The joint SEC, NASAA and FINRA investor alert warns about promoters using AI language alongside unrealistic return claims. An AI label should not substitute for checking the actual service, permissions, and evidence. No research workflow makes future returns certain.

Thesis Lab currently supports a research workflow and bounded historical simulations. Its beta does not connect brokerage order execution, and its limited historical datasets do not establish validated alpha. A saved decision is a research record, not proof that a trade occurred.

Method, sources, and limits

This guide offers an original task-evaluation framework, illustrated with invented cases. The investor warning and the current Thesis Lab release were checked on 26 October 2026. This is educational process guidance, not a security certification, suitability assessment, or recommendation to buy or sell a financial instrument. Recheck the actual application before relying on a capability claim.