The gap between an impressive demo and a system people rely on is mostly scoping. These are the five decisions we force before writing agent code, and the reason each one kills more pilots than model quality does.
Pick a workflow with an owner and a number
An agent needs a workflow that someone is accountable for and a metric that already exists. 'Support triage time' works. 'Improve productivity' does not, because nobody can tell you afterwards whether it happened.
If no one owns the metric today, you are not ready to automate the workflow. Instrument it first.
Build the evaluation set before the agent
Collect 50 to 200 real cases with the answer a competent human gave. This is the single highest-leverage artefact in the project: it turns prompt changes from taste into measurement and makes model upgrades a fifteen-minute decision.
Split it. Hold back a test set the prompt author never sees, or you will optimise straight into it.
- Sample from real traffic, including the ugly cases
- Grade with a rubric a human can apply consistently
- Calibrate any LLM judge against human grades before trusting it
Decide where a human stays in the loop
Map each action the agent can take against the cost of getting it wrong. Reversible and cheap runs unattended. Irreversible or customer-visible needs review, at least until the evaluation numbers earn the trust.
Design the review step as part of the product, not as a temporary scaffold. Reviewers need context, a fast approve path and a way to send work back.
Set guardrails and cost ceilings on day one
Scope credentials to exactly what the agent needs, validate outputs against a schema before anything acts on them, and log every tool call with inputs and outputs for audit.
Put a hard spend cap at the gateway. A retry loop with an unbounded budget is the most common way a pilot becomes an incident.
- Least-privilege tokens, rotated and scoped per agent
- Schema validation on every structured output
- Per-task and per-day cost limits enforced outside the model call
Write down the kill criteria
Before the build, agree what result means stop. A specific accuracy floor, a cost per task ceiling, an adoption threshold at week six. Pilots without kill criteria do not end, they just get quieter.
Want us to run this for you?
We do this work on client systems every week. Tell us where you are and we will say what we would do first.
Book a Call