Agent Builder

New question, new tool. No code, no consultant, no release.

Agent Builder is the wizard where your admin turns a question the team keeps asking into a tool the assistant can use — on any object in your org, including the custom ones nobody outside your company has heard of. Nothing is typed in SOQL. What it saves is a read, and it cannot save anything else.

Why it matters

The person who knows why the org is shaped that way builds the tool.

Most CRM AI disappoints because whoever configured it did not know that Status__c is abandoned, that the real pipeline lives on a custom object, or that Florida is spelled five ways. Your admin knows. Give them the wizard.

Your objects, including custom

Browse the org's actual schema — standard and custom objects, every field the running user can see. If your interesting questions live on Service_Contract__c, that is where the tools go.

Structurally safe

The wizard composes a SELECT or a FIND, with whitelisted operators, describe-checked names, and the asker's values as bind variables. There is no free-text query box. It cannot emit an update, a delete, or anything on a set of records.

A preview that cannot lie

“Try it” runs the draft as you, through the same class that will run the saved tool — so what you see in the wizard is exactly what the assistant will get, under your own permissions.

Building a tool

Five steps, about ten minutes.

  1. What it reads

    Pick the object. Choose look up records by their fields for a query, or search for words inside records for full-text — the same two shapes every packaged tool uses.

  2. Fields

    Tick the fields to return, and the one to order by — lowest or earliest first, highest or latest first. Cap the rows.

  3. Filters

    Add conditions in plain language: is · is not · contains the words · is one of · is more than · is on or after · is empty. For each value, choose whether the person asking supplies it or it is always this value. Mark a set as “required — any one of these.”

  4. Name it

    Call it, and describe it for the assistant: what should this tool do? and when should the assistant use this? That description is the prompt the model reads, and the wizard says so. Decide who may use it — everyone, or holders of a permission set — and whether it appears only on one object's record page.

  5. Preview

    Try it with real values, as yourself. Then save. It is live in the assistant immediately, listed under Settings → Tools like any packaged tool, and switchable off in one click.

OrgGenius Settings · Build a tool
Step one of the tool wizard: ‘Look up records by their fields’ selected, and the org’s objects as chips with Contract chosen.
Step one, on a real org’s schema — standard objects, custom objects, and OrgGenius’s own metadata, all one list.
Filters
The filters step: Contract End Date is on or before a value the person asking supplies, named endsOnOrBefore, with a description telling the assistant how to use it.
A condition in plain language — and the description the model itself will read, written right there.
Preview
The preview step: a sample value, four contract rows returned, and the SELECT query the wizard built with a bind variable.
Preview runs as you, and shows the query it built — a SELECT with a bind variable, because that is all the wizard can emit.

Agents

Bundle tools into a desk with its own tone.

An agent is a named set of tools plus instructions — a “Renewals desk” that only sees renewal tools and speaks to account managers, a “Support triage” agent for the service floor. People pick the agent; the boundary is unchanged, because an agent can only use tools that already exist.

  1. Name it

    A name and a one-line tagline people will see.

  2. Choose its tools

    From a searchable catalogue of every tool in the org — packaged and yours — showing the description the model itself reads.

  3. Instructions

    Guidance for the model, added after the shipped system prompt. Tone, priorities, what to ask before answering.

  4. Review

    Check it, switch it on, save.

Your own actions

The same surface registers an Autolaunched Flow you wrote as an action. Pick the object, find the Flow, name it, say when the assistant should use it, write the plain-language confirmation the person will read, mark what is required of the model, and decide who may use it.

  • It takes one record id. A Flow declaring a collection is refused — at registration and again at runtime.
  • OrgGenius verifies the person can see the record before the Flow runs. Flows run in system mode and will not do that for you.
  • Drafted, shown in full, confirmed — like every other action. The draft is held server-side.
  • A single-record update Flow is allowed. A delete Flow never is.
OrgGenius Settings · New agent
The agent wizard’s tool catalogue filtered to ‘contract’, with Read an attached file and the custom Contracts Ending Soon tool ticked.
Choosing an agent’s tools from the catalogue — packaged and admin-built side by side, with the descriptions the model reads. Scoping keeps an agent focused; it never widens access.

What it cannot do

By design, not by policy.

No free-text SOQL

There is no box to type a query in. Every clause comes from a control with a whitelist behind it.

No writes from a tool

A tool is a SELECT or a FIND. The wizard has no path that produces an insert, update or delete.

No sets

Actions take one record id. “Update all my opportunities” has no expressible form.

No privilege

Tools run as the asker, in user mode. Building one grants nobody access to anything they could not already open.

Licensing

Included. Not an add-on.

Agent Builder is in the base licence at every tier. A consultant at $150–250 an hour builds one custom tool in a day or two; your admin builds it in an afternoon without filing a ticket. Everything the tools produce — the answers, the citations, the audit log — is there for everyone.

Ask it something your team asks every week.

The beta is free, installed with you in a sandbox first, and everything is included. What we want back is the questions it gets wrong.