🔐 Now in private beta - sign up here
AI Agents let you set up Ask Double to do work on its own, on a schedule or in response to something happening in your account.
An agent is really just three things: instructions (what you want done), a trigger or schedule (when it should run), and a scope (which clients it can touch). Everything else is optional.
Building your first agent
Agents live in the left sidebar under Ask Double → Agents. The Agents list shows every agent in your practice, who created it, who last edited it, the status (Active or Paused) and when it was last run.
To create an agent:
Go to Agents
Start from a template or click +Create agent manually
If you already have agents created, click Build agent in the top right corner first.
Write your instructions — a plain-language description of the work you want done.
Add a schedule, a trigger, or both.
Set Runs on clients to pick which clients the agent applies to.
Give the agent a Name and Description (optional)
Click Save. Your agent is live and will run the next time its schedule or trigger fires.
You can also click Run at any time to test an agent without waiting for its trigger.
Writing instructions
Instructions are the most important part of an agent. Write them the way you'd write a handoff note to a new team member — specific about what to do, and explicit about what not to do. (See our templates for examples.)
Using skills in instructions
If you've already built a skill for a process — a payroll accrual, a month-end checklist, a client-specific workpaper — you can reference it directly in your instructions rather than rewriting the steps. The agent will follow the skill when it runs.
Triggers
Triggers run an agent in response to something happening in Double. An agent can have more than one trigger, and the run history shows which event kicked off each run.
Available triggers include:
Tasks — a task is created, completed, or has its status changed.
Closes — a close is created, completed, or has its status changed.
Tax projects — a project is created, completed, or has its status changed.
Files — a file is uploaded to the client uploads inbox, or attached to a task.
Client questions — a transaction or non-transaction question is added or responded to.
Client properties — a client property is updated.
Triggers respect the agent's Runs for scope. An agent scoped to three clients won't fire on a task completed for a fourth.
Schedules
Schedules run an agent at a set time, using the same recurrence options you're used to from recurring tasks.
Daily, weekly, monthly, quarterly, or yearly, with an interval (for example, every two weeks).
Specific days — a calendar day (including the last day of the month), a weekday, or a business day.
A time and time zone for the run to start.
This is what makes "on the 5th of every month, run the payroll accrual" possible as a single agent rather than a recurring reminder you have to action yourself.
Run as Combined vs. Separate
Choose how the agent runs across clients: run separately for tailored, client-by-client actions, or combine all clients into one run for a consolidated summary or report.
Use separate when the agent should make a distinct decision or take a distinct action for each client, such as:
Send each client’s contact a tailored follow-up message based on that client’s unanswered questions.
Create a client-specific custom task when a client asks a question in the portal.
Use combined when the output benefits from comparing or aggregating clients—one daily Slack digest, a prioritized queue, or a report of trends across the portfolio.
Runs for: choosing client scope
Every agent needs a Runs for setting. This controls which clients the agent's triggers and schedules apply to.
Selected clients — the agent only runs for the clients you pick. You can start with one and add more later without duplicating the agent.
All clients — the agent runs for every current and future client in your practice. This option is limited to Admins.
Scope is shown throughout the app as a label like "All clients," "Demo Co.," or "3 clients," so you can tell at a glance how wide an agent's reach is. You can also filter the Agents list by client to see everything running against a particular account.
We strongly recommend starting a new agent on a single client, reviewing a few runs, and then expanding the scope once you're happy with the output.
Reviewing runs
Open any agent to see its run history, most recent first. Each run shows what triggered it, what the agent did step by step, and what it produced.
A few things worth knowing about runs:
Agent runs are read-only. You can read the full transcript of what the agent did, but you can't reply to it mid-run or ask follow-up questions in that thread. For back-and-forth work, use a regular Ask Double chat.
Anything an agent changes is recorded in the activity log, attributed to that specific agent, so there's a clear audit trail.
Run visibility on the agent settings controls who on your team can open its runs.
Permissions and ownership
Agents are shared practice resources rather than private to the person who made them. The creator is shown for attribution, and the Agents list includes an Edited column showing who last changed the agent and when.
Admins can view, edit, and open the runs of every agent in the practice, and are the only ones who can scope an agent to all clients.
Anyone building automation that affects a client's books should expect that client's team to be able to find it.
Credits and spend
Agent runs consume AI credits like any other Ask Double work. Because agents run unattended, it's worth watching their spend for the first few weeks.
Per-agent spend is broken out under Practice settings → AI settings → Usage.
Double checks your credit balance before starting a run. If your practice is out of credits, agents are paused rather than run, and will resume once credits are available.
See AI Credits: Usage, Budgets, and Adding Credits for how credits and budgets work.
Model selection
Each agent has a model setting. Double sets a sensible default, and in most cases you should leave it alone. If you find a particular model gives better results for a specific job, you can change it per agent.
Guardrails and limits
A few boundaries exist to keep unattended work safe and predictable:
Agents cannot create other agents.
Agents can't kick off endless chains of themselves — recursive triggering is blocked.
Runs are monitored for looping behavior.
Agents can be disabled at any time from their settings, which stops all future scheduled and triggered runs without deleting the agent or its history.
Best practices
Start narrow. One client, one job, one trigger. Expand once you trust the output.
Say what not to do. Explicit limits ("don't post anything," "only act when there's exactly one match") matter more than detailed instructions on the happy path.
Ask for a report. Telling the agent to summarize what it did — including what it skipped and why — makes reviewing runs much faster.
Prefer read-and-report over write-and-hope for a new agent. Have it tell you what it would do first, then loosen the instructions once you agree with its judgment.
Reuse your skills. If a process is already documented as a skill, point the agent at it instead of restating the steps.
Check spend early. Look at per-agent usage after the first week to catch anything running more often or longer than you expected.