Salus
Guardrails to validate your agent's actions before they execute
About
Your agent processed a refund without looking up the order ID, costing you thousands. You only found out three hours later from a support ticket. Evals, output scoring, and observability can reduce the likelihood of mistakes like these occurring - but there's no solution that inspects and prevents an action as it’s about to execute. Salus does that. We’ve built an API that wraps around your agent and checks its actions at run time, blocking incorrect ones and providing immediate feedback to guide retries. Kevin and Vedant were roommates at Stanford, where they both studied computer science.
Founders
AI Research Report
Problem & Solution
Problem / Solution Report
Core Problem Addressed
Modern Large Language Model (LLM) agents are increasingly capable of taking autonomous actions, but they often lack the necessary constraints to prevent real-world harm or financial loss. Common issues include agents issuing refunds without verification, sending emails with hallucinated data, or leaking personally identifiable information (PII). Traditional observability tools are reactive, meaning they only detect these incidents after they have occurred, which is insufficient for high-stakes enterprise environments.
Salus’ Solution & Approach
Salus provides a runtime API that acts as a protective wrapper around AI agents, intercepting every tool call at the moment of execution. Unlike post-hoc evaluation tools, Salus validates actions against predefined policies before they are carried out. Key features include evidence grounding, which requires actions to be supported by prior tool outputs or conversation history, and policy enforcement through expressive, version-controlled policies written in YAML or plain English.
Self-Repair and Integration
A unique aspect of the Salus solution is its 'Self-Repair' capability. When an action is blocked, the API returns structured feedback to the agent, allowing it to understand the failure and attempt a corrected retry. Internal benchmarks suggest that approximately 58% of blocked actions can recover and complete the task correctly. The solution is designed for developers, offering a simple integration model using a Python pip package and decorators compatible with major frameworks like OpenAI, Anthropic, and LangChain.
Value Proposition
The primary value proposition of Salus is the shift from reactive monitoring to proactive prevention. By providing a systematic and testable layer of runtime checks, Salus enables enterprises to deploy agentic systems with greater confidence. This reduces the risk of catastrophic mistakes, lowers costs associated with erroneous actions, and ensures better compliance and auditability for AI-driven workflows.
Market & Competitors
Market & Competitors Report
Market Trends & Target Audience
The market for AI agent safety is driven by the rapid proliferation of autonomous automation and increasing regulatory pressure, such as the EU AI Act and NIST AI RMF. Salus targets engineering and security teams in regulated verticals like finance, healthcare, and e-commerce. These organizations require real-time reliability to prevent operational errors that could lead to legal or financial repercussions.
Competitive Landscape
The competitive landscape is divided into several segments. Traditional LLMOps and observability vendors like Fiddler AI, Arize AI, and LangSmith focus on post-deployment monitoring and evaluation. Governance players like Credo AI and Monitaur provide policy enforcement and compliance tooling. Additionally, AI security firms such as Protect AI and Enkrypt AI focus on defending against adversarial attacks and data leakage. Salus distinguishes itself by focusing specifically on the runtime interception and blocking of agent actions.
Salus’ Competitive Advantages
Salus's primary advantage is its specialized focus on the 'action-level' validation of agents, a niche that is currently less crowded than general observability. Its developer-first approach, featuring easy-to-integrate pip packages and decorators, lowers the barrier to adoption. Furthermore, its early benchmarks showing significant reductions in misalignment and cost provide a data-driven argument for its effectiveness compared to standard agent deployments.
Risks and Challenges
As an early-stage startup, Salus faces the risk of larger, better-funded incumbents in the observability or cloud provider space adding runtime checks as a feature. There is also the challenge of market fluidity; if guardrails are eventually bundled into broader security or MLOps platforms, a pure-play vendor like Salus may find monetization more difficult. Finally, with a small founding team, the company's ability to scale enterprise sales and support remains a potential constraint.
Total Addressable Market
Quantitative & TAM Report
Relevant Markets and Reported Figures
The market for Salus is situated at the intersection of several high-growth sectors. The AI Guardrails market is projected to grow from USD 0.7 billion in 2024 to USD 109.9 billion by 2034, representing a CAGR of approximately 65.8%. Simultaneously, the broader AI Agents market is estimated at USD 7.63 billion in 2025 and is expected to reach USD 182.97 billion by 2033. These figures provide the upper bounds for the potential market Salus can address.
Methodology for TAM Estimation
To estimate the Total Addressable Market (TAM) for Salus, we define its primary serviceable market as runtime guardrails for AI agents and tool-call validation. This is a specialized subset of the broader AI guardrails and agents markets. The methodology involves using the 2025 baseline for AI agents (~$7.6B) and AI guardrails (~$0.7B) as anchors. We then assume that guardrails will represent a 5-15% share of the total spend on AI agent infrastructure as enterprises prioritize safety and compliance.
Quantitative Range Estimate
In the near-term (2026-2028), the Serviceable Available Market (SAM) for runtime guardrails for agents is estimated between USD 0.5 billion and USD 1.6 billion. This is based on the projected growth of the AI agents market to roughly $11B by 2026. In the long term (2030+), if the specialized guardrails market scales toward the $100B+ projections, the TAM for companies like Salus could reach multiple tens of billions of dollars, depending on whether guardrails remain a distinct product category or become a standard feature of larger platforms.
Caveats and Assumptions
These estimates are subject to significant variance based on how 'guardrails' are defined by market researchers. Some reports include guardrails within MLOps or security budgets, while others treat them as a standalone vertical. The growth rates (CAGRs) across related reports range from 40% to 65%, and the actual TAM will depend heavily on the rate of enterprise adoption of autonomous agents in regulated industries.
Founder Analysis
Founders and Background Report
Founders & Leadership
Salus was founded in 2026 and is led by a two-person founding team: Kevin Pan (Founder / CEO) and Vedant Singh (Co-founder / CTO). The pair were former roommates at Stanford University, where they both studied computer science. They transitioned from their academic backgrounds to building Salus full-time to address the need for runtime guardrails in AI agent deployments.
Professional Experience and Roles
Kevin Pan serves as the founder driving the product and company strategy. His professional focus, as indicated by public listings, is centered on preventing AI agents from making operational mistakes during runtime. Vedant Singh, the Co-founder and CTO, brings a background as an AI researcher. Prior to Salus, he was affiliated with Stanford Computer Science, where he focused on AI research, providing the technical foundation for the company's guardrail technology.
Education and Prior Work
Both founders are alumni of Stanford University's computer science program. Their educational background is a core part of their professional identity, as highlighted in their Y Combinator (YC) company page. While detailed histories of prior corporate ventures were not extensively documented in public scrapes, their roles at Salus leverage their specialized knowledge in AI research and engineering.
Additional Notes on Team & Backing
Salus is part of the Y Combinator Winter 2026 batch, which serves as its primary institutional partner. The company maintains an engineering-first culture, evidenced by its focus on developer tools such as pip packages and code snippets. The team remains compact, with public records indicating a team size of two as of early 2026, operating out of San Francisco.
Unlock Full AI Research Report
Enter your email to access the complete analysis.
We'll never spam you. Unsubscribe anytime.