Shouldst

Inspiration

Formal method for computer programs for verification.

What is Shouldst?

Shouldst is a lightweight runtime verification layer for AI agents.\ It wraps your agent’s tools and outputs in a policy-aware safety monitor, catching unsafe or non-compliant actions before they execute and recording every decision for reproducible verification.

Think of it as a “safety kernel” for your agent runtime.

What it does

It creates policies (guardrails) and monitor trace are following all the policies.

How we built it

Before each tool call, Shouldst evaluates:

• Signals such as is_payment, contains_pii, or no_consent
• Policy verdicts from the streaming LTL monitor (G, F, U, etc.)
• Rules such as “block payments without consent,” “scrub PII,” or “warn on risky actions”

If any rule fails, the call is blocked or rewritten before it reaches the underlying system.

What's next for Shouldst

Generate policy and predicates using LLM

Framework integrations

Shouldst integrates with:

• Google ADK (python)
• Mastra (typescript)

Built With

• adk
• mastra
• python