Introducing Fitila Agents

Fitila Agents is an agent framework from Fitila Labs for scientific and engineering work.

Fitila Agents includes fifteen planning strategies. In the hypothesis-and-evidence planner, a question is framed as competing hypotheses; the language model selects which evidence to collect, and an evidential core combines the observations using Dempster–Shafer theory. Dependent sources count once, and evidence used to find a hypothesis is not used to confirm it. The result is a verdict with belief and plausibility for each hypothesis, which can be inconclusive, and a ranking of the next measurements by expected value per unit of cost. A model-discovery planner fits competing model structures and checks whether their parameters can be identified. A hierarchical task network planner runs an established procedure step by step, with each step bound to a tool.

The tools are our scientific libraries. Over the Model Context Protocol, agents call our numerical library (optimization, differential equations, statistics, control, uncertainty), our mesher (meshes from a declarative model language) and our quantitative-biology library, and report the results those libraries return.

In Agent Studio, teams build agents on a canvas, run them and watch each tool call, test them against evaluation suites, and publish fixed versions stamped with the runtime they were built against. Each tool's access is a permission, spending is capped, and actions that change data or systems wait for human approval.

The planners, the evidential core and Agent Studio are described on the product page. Fitila Agents →