Best practices for creating custom Agents with Agent Builder

A well-designed custom Agent captures the steps, logic, and decisions behind a workflow so it executes the same way every time. The quality of an Agent depends on how clearly you define its purpose, behavior, and instructions. Use the guidance below to build Agents that produce consistent, high-quality, traceable outputs.

Pre-requisites:

• DataSnipper version 26.1 or later
• Accelerate or Elevate DataSnipper package
• Access to the internet

1. Start with a workflow you already do well

The best custom Agents come from procedures you have already refined manually. Before opening Agent Builder, make sure you can answer:

• What is the end-to-end procedure, from input to deliverable?
• Which documents or data does it rely on?
• What decisions or judgements are made along the way?
• What does a finished, reviewable output look like?

If a process is still evolving, document it first, then build the Agent.

2. Write a clear Name and Description

The Name and Description help your team understand at a glance what the Agent does and when to use it.

• Name: Use a specific, action-oriented title (e.g., "Payroll Reconciliation Agent", "Revenue Cut-off Testing Agent"). Avoid generic names like "Audit Helper".
• Description: In one or two sentences, explain the procedure the Agent performs and the type of workpaper it is designed for. This is what teammates will see when selecting an Agent in Excel.

3. Define a single, specific Goal

The Goal tells the Agent what success looks like. Keep it focused on one outcome.

Do:

• "Validate payroll accuracy and summarize key discrepancies above €500."
• "Reconcile bank statement transactions to the general ledger and flag unmatched items."

Avoid:

• Combining multiple unrelated procedures into one Agent.
• Vague goals like "Review the data and tell me what looks wrong."

If your workflow has two distinct objectives, build two Agents.

4. Use Behavior to share the context only your team knows

The Agent already understands how to perform standard actions like snipping, extracting, and writing back to Excel, those mechanics are handled for you. The Behavior section is where you share the context a new joiner would need on their first day at their desk: the nuances of your methodology, what "good" looks like in your firm, and the situations where judgement is required.

Think of it as briefing a junior auditor, not configuring software.

Useful context to include:

• Methodology and policy choices. "Materiality threshold for this engagement is €500." "Treat related-party balances separately from third-party balances."
• Definitions specific to your firm or client. "A 'cut-off issue' means an invoice dated in December but recorded in January."
• Edge cases and exceptions. "Loans denominated in a non-EUR currency should be flagged for review, not converted."
• When to stop and ask. "If a loan agreement has an addendum that changes the interest rate, treat the addendum as the source of truth."
• What the deliverable should feel like. "Findings should be written for a manager review, concise, with references to source documents."

You do not need to tell the Agent to "always check carefully" or "be accurate", that is already baked in. Focus on what is specific to your workflow that someone new would not know.

5. Write Instructions that describe the work, not the workbook

The Instructions section is the procedure itself. The most common mistake here is describing the workbook ("compare column D to cell B25") instead of describing the work ("the total should reconcile to the invoice amount").

When you describe the work in business terms, the Agent can apply the logic correctly even if the workbook layout changes. When you describe the workbook, the Agent becomes brittle and only works for one specific template.

Describe the work like this:

• "The total of the amortisation schedule should reconcile to the trial balance interest expense."
• "Each invoice in the sample should be agreed back to the supporting bank statement for the same amount."
• "Recalculate interest using the rate and day convention stated in the loan agreement."

Avoid describing the workbook like this:

• "Check that column D equals cell B25."
• "Look at row 12 and compare it to row 18."
• "Put the result in cell F4."

If a step requires judgement, say so explicitly and ask the Agent to flag it for human review.

6. Test before you share

Always run the Agent on a sample workpaper before changing its availability more widely.

• Run it on a representative file with known answers.
• Review the output for accuracy, completeness, and tone.
• Refine the Goals, Behavior, or Instructions if the output is inconsistent.
• Test edge cases (missing data, mismatched columns, unusual values) to see how the Agent handles them.

7. Publish and share

Once you are satisfied with the Agent, change the Status to Published so it is ready for use.

Sharing an Agent with the wider team or organisation depends on the role assigned to you in DataSnipper:

• Content Manager — can publish Agents firm-wide.
• Agent Builder User — can build and run Agents.

Who can assign these roles

The Dashboard Owner is the role responsible for assigning Agent Builder roles. Dashboard Owners can do this themselves — no DataSnipper-side action is required.

To assign a role:

1. Go to Users.
2. Click the user you want to update.
3. Assign Content Manager or Agent Builder User as appropriate.

If you would like an Agent shared firm-wide and you are not a Content Manager, speak with your Dashboard Owner.

8. Keep Agents maintained

Custom Agents reflect your methodology, and methodologies evolve. Plan to:

• Review published Agents periodically, especially after methodology updates.
• Update Instructions when audit standards, thresholds, or templates change.
• Retire or replace Agents that are no longer aligned with current practice.

Quick checklist before publishing

• The Name and Description make the Agent's purpose clear at a glance.
• The Goal describes a single, specific outcome.
• Behavior includes guardrails that protect source data and require human approval where needed.
• Instructions are ordered, concrete, and include thresholds and output format.
• The Agent has been tested on a representative sample and produces consistent results.
• Visibility is set appropriately for the Agent's maturity.