Skip to main content
An Agent is the thing that prepares the work. You either create one directly (Advanced) or get one automatically by choosing a task template (Quick). This page covers the direct route and the guided builder.
Acct Agent Builder

​
The six-step builder

Creating an agent walks through six steps — the same shape whether you’re starting fresh or refining a template’s agent.
1

Basics

Give the agent a name (e.g. “R&D Tax Credit”) and a description of what it does for a workstream. The name generates a fixed identifier (e.g. r_d_tax_credit) that can’t change after creation, so choose it deliberately. An Enabled toggle controls whether it’s visible to workstreams.
2

Capabilities

Pick the skills and plugins this agent may use — Read documents, Query Data Store, Generate workpaper, Web search, Tax Forms, and so on. Turn on only what the work needs; a skill grants capability, not access. See Instructions and skills.
3

Instructions

Write plain-language direction for how the agent should behave and sequence its work — the glue logic. Be specific about the output you want and any standards to follow.
4

Workstreams

Choose which workstreams (clients) this agent runs for. Agents bind to their workstreams here, at build time.
5

Policies

Set the guardrails — the constraints the agent must operate within (for example, where human review is required). This is where review and escalation behaviour is shaped.
6

Review

Confirm the configuration and create the agent.

​
What naming an agent does

Creating an agent silently provisions a Space — an isolated bucket for that agent’s data. You don’t manage the Space directly; you work in terms of the agent. It’s simply why the agent’s working data stays separate from everything else. See Agents, Spaces, and Skills.

​
Enabling and testing

The Enabled toggle controls whether workstreams can see and use the agent. A common pattern is to leave it disabled, test it on a client with real files, and enable it only once the output looks right.
A created agent isn’t useful until it has instructions and skills, access to data (a skill alone doesn’t grant it — see Setting Doc Store access), and a successful test run.

​
Next steps

  1. Instructions and skills
  2. Setting Doc Store access
  3. Testing an agent → Schedule and publish