Salus (YC W26)

guardrails to validate your agent's actions before they execute.
Backed by Y Combinator.

[Y Combinator @ycombinator]

.@salus_ai is building runtime guardrails to validate your AI agent’s actions before they execute.

By adding their API to your codebase, incorrect actions are automatically blocked and your agent receives actionable feedback to self-correct.

Congrats @thevedants1 and @pankev18 on the launch!

https://t.co/YnbuVT2xzs

[Salus (YC W26) @salus_ai]

Appreciate that.

This is exactly our thesis. The smaller or more local the model, the more important runtime enforcement becomes.

In our internal benchmark testing with Salus, we’ve seen smaller models show substantial gains in effective reliability once runtime validation is enforced, while preserving their cost advantage.

We’re model-agnostic by design. We already integrate with frameworks like LangChain and CrewAI, and support for additional local runtimes like the ones you mentioned is in progress.

Would be interesting to see this applied to LM Studio or Ollama setups!

[Salus (YC W26) @salus_ai]

Exactly.

That boundary between reasoning and execution is where most failures hide.

You can have correct reasoning and still trigger the wrong side effect.

That’s why we believe runtime enforcement isn’t optional, it’s the control layer that turns agents from “best effort” systems into reliable ones that can actually be used by developers.

[Salus (YC W26) @salus_ai]

Great question.

We don’t evaluate actions in isolation. Each action is validated against prior tool outputs, required prerequisites, and accumulated policy state across the run.

We model execution as a stateful graph, not a sequence of independent calls.

If something looks invalid locally but is justified by earlier context, that prior state is part of the evaluation before we decide to block.

Curious what edge cases you’ve run into there?

[Salus (YC W26) @salus_ai]

A lot of agent failures get labeled as “hallucinations.”

In practice, we have seen two cases:

The model reasons incorrectly.

The model reasons correctly but executes an invalid action.

We intercept both at runtime.

Before any action executes, we validate it against:
• Required tool prerequisites
• Accumulated execution state and evidence
• Explicit policy invariants

If it violates constraints, we stop it before it fires and return structured feedback so the agent can self-correct.

Stopping incorrect actions before they execute is the difference between monitoring and protection.

If you’re running agents in production, how are you enforcing runtime guarantees today?