Multi-agent Harness Bench
Five harnesses, three model pairs, 15 minutes to port a legacy app.
Software Research Engineer
Dionizije Fa
Not somebody who woke up a loser.
Latest Research Topics
The Fragility of AI Safety: How Model Guardrails Weaken Cyber Resilience
How guardrails can obstruct defense and weaken system resilience.
Beyond Benchmark Scores: An Analysis of Bioinformatics Agent Evaluations
Tradeoffs in bioinformatics evals and how to mitigate them.
Harness Bench
Benchmarking popular agent harnesses.
Sound agentic science requires adversarial experiments
The only real verifiers are experiments.
BioAgent Bench
Evaluation suite with adversarial testing for end-to-end bioinformatics tasks.