CAESAR
An agentic workflow evaluation and reliability platform combining generated tests, trace analysis and explicit safety gates.
Agentic AI
Evaluation
Research
CAESAR explores how to make agentic workflows testable, observable and accountable. I architected and built it as my MSc dissertation at the University of Bath.
The challenge
Agentic workflows make sequences of decisions and tool calls. Evaluating only the final answer leaves important behaviour unexamined. CAESAR brings test generation, execution evidence and evaluation into a single workflow.
What I built
- Pre-flight test generation for agentic workflows.
- Trace analysis to inspect execution behaviour.
- Hard safety gates alongside targeted LLM-as-judge evaluation.
- Auditable experiment provenance linking evaluation results to the underlying runs.
Research and maturity
CAESAR is an applied research platform developed for an MSc dissertation. Its focus is the evaluation and reliability of agentic workflows.