Shrink AI agents: code does the work, the model makes one narrow judgment

Flywire, a financial company, keeps its AI agents small. Ordinary code does the predictable work. The model gets one narrow step where judgment is needed. Each step's access list is written down and reviewed like other code. Eduardo Simon and Javier Casal of Flywire described this setup on the AI Engineer show.

What was said

They presented Orbit, Flywire's internal platform for engineering workflows. A speaker on the show said the team turns repeated procedural logic into code. In their words, "We call a model only where reasoning is actually needed."

The code review agent shows how it works. An engineer mentions the review bot on a merge request, which is a proposed code change. Fixed steps then fetch the details, rate the risk, load earlier findings and copy the repository. Only after that does the model run. It can read the code but cannot change it, and it must answer in a strict format. Another fixed step posts the findings.

Each agent is a YAML file, a plain settings file, stored in the repository. Colleagues check a change to that file in the same review process they use for software changes. For every step, the file limits the network, the tools, the skills and the outside servers it may use. A speaker said the access is "visible and reviewable like any other code." When a model reply does not match the expected format, Orbit tries again or takes a backup path.

Why it matters

For buyers, the lesson is a question to ask any vendor: what can each step of this agent access, and can we review that list? For builders, it means reusing review habits the team already has. Flywire also said it built in-house partly to keep sensitive data inside its own infrastructure and to skip long vendor security reviews. Not every company can make that trade.

The other side

The speakers did not pitch this as a safety fix. They spoke of reliability, control and cost. The talk showed no data on whether the design reduced incidents.

They were also open about gaps. One said the setup is not full sandboxing yet, and the team is working toward it. Another said new agents usually reach production after an informal check, and online evaluations are not yet in place.

Because Flywire is regulated, people still review and approve the merge requests the agents produce. So humans remain a key safeguard. The talk gave no figures on how many errors that review catches. It also did not say whether smaller firms could afford to build a platform like this.