AgentRunway – Trace-based invariant checker for AI agents

Provides TLA+-style invariant checks against recorded AI agent runs for auditability

Problem

Developers running AI agents in production often cannot confirm whether an agent violated key invariants under retries, crashes, or concurrency. Existing tools either require hand-writing complex formal specs or only offer after-the-fact log viewing, lacking a lightweight way to automatically check recorded traces and pinpoint violations to specific steps.

Solution

Import recorded agent runs (JSON/event streams), use preset invariant templates for common agent frameworks (e.g., 'must decrement inventory before payment', 'retries must not double-charge'), and automatically check traces for violations, highlighting the offending step, input, and timestamp. Export minimal reproducible cases for regression testing after fixes.

Users

Backend engineers and platform teams using AI agents who care about reliability, idempotency, and concurrency safety

MVP

Import event logs and visualize as a timeline; built-in 5–8 common invariant templates (payments, inventory, retries, locks); automatic scan with highlighted violation steps; export minimal repro cases; support custom simple rules via JSON expressions

Revenue

Open-source core + subscription for hosted scanning service (billed by scan count or log volume); enterprise tier offers on-prem deployment and custom rule libraries

Why now

AI agents are moving from experiments to production, exposing data inconsistency issues from retries and crashes; lightweight formal methods like TLA+ are gaining traction on HN and beyond, but automated trace-checking tools for agents are missing