Ashr

Ashr was founded in late 2025 by Shreyas Kaps and Rohan Kulkarni, two engineers with a deep history in agentic systems. Shreyas brought two years of intensive experience building AI agents specifically for high-stakes finance and devops environments. Rohan, a Berkeley EECS graduate, was previously the Co-founder and CTO of Ask Geri, a workforce management platform that was successfully acquired.

The team’s combined expertise in both autonomous infrastructure and workforce management led them to join the Y Combinator Winter 2026 (W26) batch. Based in San Francisco, Ashr is built on a philosophy of "Data-Centric Evals," where the founders believe that the only way to achieve AGI-level reliability is through rigorous, automated testing that mirrors the unpredictability of human users.

Ashr is built specifically for AI Engineering Teams and DevOps Professionals who are responsible for deploying autonomous agents into production. It is the ideal tool for companies building AI customer support agents, automated sales representatives, or "agentic" developers that need to interact with external APIs and databases.

It is also a favorite for Product Managers who need to ensure that their AI features align with business goals and safety standards. In 2026, it has become a sanctuary for QA Teams who are transitioning from manual script-testing to the more complex world of AI evaluation, providing them with a "no-code" way to generate and monitor thousands of test cases.

Ashr acts as a synthetic user generator. Its primary capability is creating large-scale "User Stories" that interact with your AI agent's tool calls and decision-making logic. It doesn't just generate one-off tests; it maps out entire branching journeys, picking up on errors, inconsistencies, and failures that would otherwise go unnoticed until a customer encounters them.

One of its most powerful features is Custom Evaluation Metrics. You can define specific business goals—such as "Never offer a refund over $50" or "Always verify the user's ID before changing a password"—and Ashr will automatically grade your agent’s performance against these rules across every simulated journey. It provides a real-time "Health Score" for your AI agents, allowing for a data-driven approach to fine-tuning and prompt engineering.

As a YC W26 company, Ashr’s pricing is primarily Usage-Based, following the "Consumption" model that has become standard for developer tools in 2026. While they offer a Free Sandbox for developers to test their first few agents, scaling into production testing involves a "Pay-per-Journey" fee. This ensures that startups only pay for the amount of testing they actually need during their development cycles.

For enterprise organizations, Ashr provides Custom Annual Plans that include dedicated support, on-premise deployment options for sensitive data, and higher throughput for massive agent fleets. Because the platform is still in its high-growth phase, most pricing is finalized after a consultative pilot to determine the volume of tool-calls and evaluations required.

The most critical thing to watch on Ashr is Calibration Accuracy. Because Ashr "mimics" users, the quality of its evaluation is only as good as the "User Archetypes" you define. If you don't provide enough variety in the user types, the testing may miss specific edge cases. Additionally, pay attention to Integration Latency. Running thousands of simulated journeys through your agent can be resource-intensive, so ensure your dev/staging environments are configured to handle the high volume of traffic during a test run.

In the niche of AI agent testing, the primary alternatives are LangSmith (LangChain) and Braintrust. LangSmith is excellent for developers already using the LangChain framework, while Braintrust is a heavy-hitter in the enterprise space for general LLM evaluations. Promptfoo is a popular open-source alternative for developers who want to run basic evals locally without a SaaS subscription. Ashr distinguishes itself by focusing specifically on the "agentic" side—mimicking tool calls and multi-step journeys rather than just single-turn prompts.