Krv Labs provides evaluation infrastructure designed for agentic systems, focusing on delivering comprehensive evidence packages rather than simple pass/fail scores. The platform addresses the need for safe and compliant deployment of AI agents by justifying every agent decision and highlighting areas of uncertainty. Its approach centers on assembling proof from three critical dimensions: the data an agent relies on, the internal reasoning within the model, and the code generated by the agent.
Key features include Pulsar, which analyzes the structure of datasets, validates new data instances, and identifies regions with insufficient coverage or clusters of failures. Pasteur, currently in stealth, is responsible for stress-testing models under extreme conditions, exposing hidden failure modes by probing internal reasoning. Topos, in development, audits AI-generated software by examining its structure and logic, detecting security vulnerabilities and documenting the soundness of code beyond basic execution checks. The platform's verification process is mathematically grounded, evaluating the underlying structure of agent answers, the boundaries where behavior may become unsafe, and the proof obligations embedded in code.
Krv Labs is intended for organizations and teams deploying agentic systems who require auditable, transparent, and defensible evidence of agent behavior. Its evidence packages are designed to assist with regulatory compliance, operational safety, and internal or external audits. The platform also aims to prevent hallucinations by detecting when an agent operates on data it does not fully understand, and to mitigate AI code debt by automatically auditing agent-generated software for security and structural integrity.
The core verification libraries powering Krv Labs are open source and released under the BSD 3-Clause License, supporting transparency and reproducibility. These libraries include Pulsar for topological data analysis, Topos for program graph analysis across multiple programming languages, Phil for representation-guided imputation of missing tabular data, and Trailed for structure-preserving representation learning. Krv Labs' infrastructure is built in public, emphasizing transparency in AI safety and reliability.
In the LLM eval & observability space, Krv Labs takes a focused approach. It focuses on evaluating and justifying the behavior of AI agents to ensure safe and compliant deployment. Krv Labs is a B2B product aimed at AI developers and organizations deploying agentic systems. It runs on the web.
Behind Krv Labs is Krv Labs, and the product first shipped in 2026. The project is developed in the open on GitHub with 73 commits in the last 90 days. Among its 6 catalogued features are agent evaluation, evidence packages, and data inspection.
Latest indexed changes and source events
Module decomposition cut agent token use 32% on follow-up feature additions verified by the PulseGate indexer
Other apps tracked under the same category.
