Skip to main content
An AI agent connected to toorow reads governed marketing data. Five rules keep its answers correct; the guarantees below are what the platform holds up on its side.

Five rules for an agent

Rule 1 — Discover before you query

Inspect what a source actually offers with get_source_capabilities(project_id, connection_ref_id) before requesting a metric/dimension combination. Never guess a column name or a grain: a guessed field does not fail loudly, it returns nothing.

Rule 2 — Never sum a ratio

CTR, CPA, ROAS and Average Position are not additive. Summing them across days or campaigns produces a number that looks plausible and is wrong. Query the semantic ratio views instead — semantic_ctr, semantic_cpa, semantic_roas, semantic_avg_position — which recompute the ratio at the grain you asked for.

Rule 3 — Always scope to a project

Every tool call carries an explicit project_id. There is no cross-project read: the isolation is enforced by the database and by the warehouse, not by convention, so an unscoped call fails rather than leaking.

Rule 4 — Keep structuredContent intact

Answer with a short text summary (≤ 30 lines) and pass the full structuredContent payload through untouched. The host renders interactive widgets from that payload; a summary that replaces it turns a chart into a paragraph.

Rule 5 — Context events are candidates, never causes

When annotating an anomaly with context_events, present them as possible context. Causal phrasing is prohibited — “caused by”, “due to”, “because of”. toorow records that a deployment and a traffic drop share a date; it does not know that one produced the other.

What the platform guarantees


Next Steps & Cross-References

Agent Tools Reference

Inspect FastMCP tool definitions and parameter schemas.

Platform Constraints

Review security rules and pre-live checklists.