AgentBench
Pass/fail gates on agent output. Property oracles on pull requests — tests hold, protected files stay untouched, scope doesn't drift. Deterministic. Not a score on a dashboard.
Software is starting to write itself.
Casualstack is building the infrastructure that makes autonomous systems accountable — before they touch production, not after something breaks.
Scroll
What we're building
Intelligence without accountability is just automation with better marketing.
Models got good fast. Agents ship code, configs, and decisions into real systems — often with nobody checking whether the output still matches intent.
Dashboards score traces after the fact. We think the next layer of infrastructure looks different: hard gates, replayable execution, cryptographic proof — the same rigor production software has had for decades, rebuilt for systems that think.
We started with developer tools. We won't stop there. Any field where autonomous software touches the real world will need this layer.
Casualstack is early. The work is serious.
The stack
Three primitives. One direction.
Witness
Causal replay and bisect for agent runs. Find the exact moment execution diverged — not a timeline to stare at, a needle to pull.
K8sAttest
Proof-carrying infrastructure deploys. Agent-generated configs with signed least-privilege attestation — correct before they hit the cluster.
Built for what's coming, not what's comfortable.
Coding agents are the wedge. The problem is bigger — any autonomous system that modifies state needs a trust boundary. Finance, operations, infrastructure, research. Different surfaces. Same requirement: you should be able to prove what happened, block what shouldn't, and replay what did.
We're hiring the future into existence.
Follow the work. First releases land on GitHub. If you're building where agents meet production, you already know why this matters.