Join our growing team. See the job board →
Talk to us →

What an AI agent needs before it ships

A working demo is not a finished agent. Here are the four things we check, from a clear task to a way to know when it breaks, before an agent goes in front of real users.

By Faisal Kawar · · 2 min read

A central agent box connected by lines to seven smaller agent boxes on a green grid

An agent that works in a demo can still fail on a Tuesday afternoon with a real customer. The gap between the two is not magic. It is a short list of things that are easy to skip when a demo goes well.

These are the four we check before an agent is allowed near real users.

1. One workflow, described plainly#

Agents do best when they own one workflow. "Handle support" is too wide. "Read an incoming refund request, check it against the policy, and draft a reply" is a job you can test.

Write the job the way the person who does it today would describe it. If two colleagues would disagree about what a good result looks like, the agent will not settle it for you.

2. A pass mark before you build#

Decide what "good enough" means first, with examples. Pick twenty real cases, including the awkward ones, and write down what a correct answer looks like for each.

This list becomes your test set. It is also the quickest way to see whether a change made things better or worse, which is why we treat it as part of the build and not an afterthought. Our evaluation service is built around exactly this step.

3. A person in the loop where it matters#

Some actions are cheap to undo and some are not. Sending a draft is easy to review. Moving money or deleting records is not.

  • Let the agent act alone where mistakes are small and reversible
  • Ask a person to approve anything that is costly or hard to undo
  • Record every decision, so a reviewer can see what the agent did and why

4. A way to notice when it breaks#

Models change, data drifts, and customers ask new things. Without monitoring you find out from a complaint.

Keep sampling real conversations after launch, and send a small share to expert reviewers each week. Their notes tell you what to fix next, and they give you fresh cases for the test set.

The short version#

Pick one workflow, set the pass mark, keep a person where the stakes are high, and keep watching after launch. Do those four and the agent that shipped is the one you tested.

If you are planning an agent and want a second pair of eyes, talk to us. We build agents and we test them the same way we test models.

← All postsView as Markdown

Keep reading

A target with green hit marks next to four bars measured against a pass line on a green grid
· 2 min read

How to evaluate an LLM beyond the benchmark

Public benchmarks tell you how a model does on someone else's questions. Here is how to build tests from your own work, probe for safety gaps, and decide when a model is ready to ship.

Two document outlines side by side on a green grid, the right one highlighted with a check mark
· 2 min read

Why human feedback still matters for fine-tuning

Models learn from examples, but what counts as a better answer is a human judgement. A look at how expert comparisons turn taste and domain knowledge into a training signal.

Want to talk about your AI project?

Talk to us →See our work →