Skip to content
Fitila Labs

Research · 03 AI

Agentic science

  • Liveframework and Agent Studio
  • In buildproduct agents

Our agents report the queries they ran, the numbers they computed, and the approvals they requested before changing anything. They use the same data gateway and scientific libraries as the rest of our software.

Current work

  • Fitila Agents, a multi-provider agent framework with fifteen planning strategies. They include a hypothesis-and-evidence planner whose belief comes from a Dempster–Shafer evidential core, a model-discovery planner, and hierarchical task network, belief-state and graphical-model planners. Around them sit guardrails, budgets, a hash-chained evidence ledger and human approval gates.
  • Agent Studio, a browser canvas to build, test, compare, evaluate and publish agents, that records every tool call. Every option it shows comes from what the installed runtime can actually build.
  • Scientific tools over MCP. Agents call the same numerical, biology and meshing libraries our products use, and report the results they returned.
  • Tiered product agents. In TerraNavitas, agents escalate by autonomy: read, then compute, then plan, then act with approval.
Agent Studio paused at an approval gate, showing the evidence behind its proposal
Screen: Agent Studio waiting for a person to approve hit finding in a showcase workflow. Every number is computed; no model is asked at any step.
planning strategies, including hierarchical task networks and hypothesis–evidence
15planning strategies, including hierarchical task networks and hypothesis–evidence
built-in tools, each tagged with what it can reach
96built-in tools, each tagged with what it can reach
numbers behind every verdict, from belief to conflict
6numbers behind every verdict, from belief to conflict

Open problems

  • Evaluating agents. Scores from model judges drift, and labeling task outcomes is expensive. We are building evaluation suites that check the evidence an agent cites as well as its answer.
  • Provenance across agents. When one agent delegates to another, the record of which agent made each decision has to be preserved.
  • Calibrating evidence packs. Our evidence packs encode each field's practices as data. Their parameters are defaults until practitioners in each field review them.

The framework and its studio are described in full on the Fitila Agents page. Fitila Agents →