Kelter: an AI agent engine that withstands model mistakes

In one sentence

Kelter is a runtime for AI agents in which the model suggests the next step of the process, while a predictable program runs the work. This architecture makes the system resilient to model errors.

The problem

An AI agent is a program that carries out a task step by step on its own: it looks things up, uses tools and saves results. In most of today's solutions the language model steers the whole process. When it gets something wrong, the process falls apart.

So companies pick the largest and most expensive models, because only they make mistakes rarely enough to rely on. On top of that, every new feature often means rebuilding the application.

The idea

I reversed the roles. The model proposes what to do next. The program around it checks the proposal, carries it out and keeps order: it retries failed steps, records every event and blocks any action outside an agreed list.

For the business, this means the same process runs on a large cloud model and on a small model on your own server. The model determines the quality of decisions. Whether the process finishes cleanly depends on the program.

What already works

Kelter runs as a web app, a desktop app and a container service. It can hand tasks to other agents and accept tasks from them, lay out work as a plan of steps and calculate the cost of every run. What remains is signing the desktop app and deploying to the target server. The repository is private for now.

Compared with other tools

The individual parts of Kelter exist in other tools. What is new is the combination: a system that treats the model as a replaceable part and keeps the process in order even on a small, cheap model. Kelter is a working, verifiable proof of that thesis. It does not try to compete with large platforms on the breadth of integrations.

Is it a finished tool

Not for other companies yet. It has one author, no external users and no independent security review. As an open experiment it is ready: every promise can be checked with a script, and the design decisions are written down.

Kelter will be used in Exocortex R&D. After that first real use I want to open it to the public, with the lab experience as evidence.

Why it belongs in the portfolio

Kelter shows the same thing as the lab and Gromatik: the reliability of a system with AI should come from a mechanism you can check, not from hoping the model will not make a mistake. In a conversation about a leadership role, it is evidence that I talk to engineers about agents at the level of architecture.

My role

I came up with, designed and built Kelter myself, working with AI agents. It is my intellectual property.