The operating system for autonomous agents.
Agents need more than scripts. They need a formally verified environment where they build and evolve their own capabilities.
The Vision
Agents Build Tools. Temper Makes Them Safe.
Agents are starting to synthesize tools at runtime. The infrastructure to make this safe is developing. Temper connects formal verification, Cedar authorization, and evolution feedback into one framework.
| Tools | Stateless, ad-hoc synthesizers | Verified state machines that persist as specs |
| Security | Hardcoded, manual policies | Policies derived from behavioral intent |
| Growth | Failures are discarded logs | Unmet intents feed the evolution engine |
Declarative Intent.
Everything in Temper starts as a specification. Agents describe what they need — state machines, data models, and policies — and the kernel derives the runtime behavior.
If a transition is not in the spec, it cannot happen. If a policy is not in the store, it is denied. The kernel interprets intent directly.
Verification Cascade.
Before a spec is deployed, it must survive four levels of mathematical proof. We prove correctness across every possible execution trace.
The model checker verifies the actual Rust code that runs in production. If you can't prove it, you can't deploy it.
Cedar Authorization.
Every action flows through a Cedar authorization engine. Temper operates on a default-deny posture. Access is never assumed.
Denied actions surface as pending decisions. Approve once, and Temper generates the policy that governs the agent forever.
No permitted policy found.
Closing the Loop.
Failures in Temper aren't errors — they are training data. Every denied action is recorded as a trajectory entry.
The Evolution Engine analyzes these patterns and proposes spec updates. You approve the verified fix, and the agent OS grows.
Verified Shared State.
Agents coordinate through a shared state layer queryable via OData. Every entity is a verified state machine, ensuring coordination remains safe.
Background executors claim tasks, agents spawn children with scoped permissions, and everyone operates on the same event-sourced truth.
Roadmap
Built Bottom-Up.
From the kernel to full agent execution.
Persistent State
Verified entities queryable via OData API. Event-sourced truth instead of JSON blobs.
WASM Connectivity
Streaming-capable integrations as sandboxed modules mediated by Cedar policy.
Agent Execution
Headless executors claiming agent entities. Spawning and shared-state orchestration.
The living harness is here.
Join us in building the next era of autonomous engineering.
Get Started on GitHub