Skip to content
Marivo
Marivo

A Data Analysis Harness for AI Agents

A Data Analysis Harness for AI Agents
Install Marivo View on GitHub

Analyze a business question with an agent

Section titled “Analyze a business question with an agent”

Marivo is a Python analysis harness for AI agents. It gives you and an agent one reviewable loop for moving from an approved business question to a conclusion supported by typed results and recorded evidence.

For example, suppose you need to decide how to respond to a Q4 revenue decline. You provide the decision to support, the approved meaning of revenue, the periods and populations to include or exclude, and the breakdowns that are acceptable. The agent reads the semantic catalog, resolves and validates the selected inputs, performs typed analysis, and records the evidence. You then review the metric contract and accept or reject the conclusion.

What you decide and what the agent does

Section titled “What you decide and what the agent does”
  1. You set the business contract. State the decision, approve the metric’s meaning, define the scope and exclusions, and approve meaningful breakdowns.
  2. The agent executes a bounded investigation. It reads the catalog, validates the selected inputs, runs typed analysis, and preserves the material results and evidence.
  3. You review the handoff. Confirm that the metric, scope, comparison, and breakdowns still match your intent before accepting the conclusion or asking for a deeper investigation.

Four safeguards in the workflow

Section titled “Four safeguards in the workflow”

Semantic Layer

Python declarations persist datasource bindings, approved business meaning, guardrails, and refs for metrics and dimensions.

Typed Analysis DSL

Typed operators accept catalog refs and return explicit frames instead of hiding analytical steps in an unreviewable answer.

Analysis Session

A project-local session keeps the question, frames, artifacts, and investigation record together for one investigation.

Evidence Engine

Typed findings and bounded digests connect operator results to their inputs without replacing agent judgment.

From question to reviewed conclusion

Section titled “From question to reviewed conclusion”

The collaboration loop is business decision → approved metric and scope → live catalog validation → typed analysis → evidence → user review. If a required semantic object is missing or disputed, or the requested analysis exceeds the approved metric contract, the agent stops instead of silently redefining the question. New or changed semantic objects complete readiness certification in their authoring flow.

Start here

Section titled “Start here”