Skip to main content
Open a table and go to Agent tools. Everything there is generated from the columns you declared — nothing is hand-written, and nothing is editable. Copy a tool and paste it into an agent’s tools array. A table no agent references is not reachable by any agent, so this step is the whole handoff.

​
What gets generated

​
The lookup tool

lookup_<table> fetches one row by its key.
The parameter is named after your key column — plan_id, not row_key — so the model sees something meaningful.

​
The search tool

search_<table> filters, sorts and returns several rows. It gets one parameter per declared column, plus _min/_max bounds for numbers and dates, plus a limit.
Every filter is optional. The model fills in what the caller mentioned; the rest disappear. Asking for one constraint gives a one-constraint search, not a search with a handful of empty equalities that match nothing. Your descriptions and allowed values come through onto the parameters, which is why filling them in on the schema matters — that text is what the model reads when deciding how to call the tool.

​
Stored query tools

Any stored query that is validated and switched on also appears here, as its own tool. Same tab, because this page is meant to be the complete answer to “what can an agent do with this table?”. Queries that are drafts, broken, or need revalidating do not appear — handing an agent a tool that fails mid-call is worse than not having it.

​
Giving them to an agent

Copy one tool, or Copy all, and paste into the agent’s tool configuration. They’re ordinary tools — they work on voice, chat and email agents alike.
Tools are copied, not linked. Editing the table’s schema later does not change any agent that already has them. That’s deliberate: nothing silently alters a live agent’s abilities. When you want an agent to pick up a schema change, come back here and copy the regenerated tool.

​
Writing the prompt around them

The tools describe what they do. Your prompt decides when.
Worth saying explicitly in the prompt:
  • Which tool for which question. Models reach for search when lookup is exact and cheaper.
  • What not to say. Rows an agent can read but shouldn’t quote — internal margin, a discontinued flag.
  • What to do when nothing matches. An empty result is a normal answer, and without guidance a model may invent something.

​
When a lookup returns nothing

An agent gets a structured answer, not an error:
That’s on purpose. An agent mid-call can recover from “no matching records” and cannot recover from a tool that threw. The same applies if the lookup couldn’t complete — the agent is told the lookup isn’t available right now rather than the turn ending.

​
Testing

The Rows tab has the same filters the search tool uses. If a filter returns what you expect there, the agent’s tool returns the same thing — it’s the same query underneath. Then test with the agent itself: ask the question a caller would ask, in the words a caller would use, and check it picks the right tool. Most failures at this stage are prompt or description problems, not data problems.