Show HN: Kybernis – Prevent AI agents from executing the same action twice
Because agents retry steps, re-plan tasks, and run asynchronously, the same action can sometimes execute more than once.
In production systems this can cause duplicate payouts, repeated mutations, or inconsistent state.
Kybernis is a reliability layer that sits at the execution boundary of agent systems.
When an agent calls a tool:
1. execution intent is captured 2. the action is recorded in an execution ledger 3. idempotency guarantees are attached 4. the mutation commits exactly once
Retries become safe.
Kybernis is framework-neutral and works with agent frameworks like LangGraph, AutoGen, CrewAI, or custom systems.
I built this after repeatedly seeing reliability failures when AI agents interacted with production APIs.
Would love feedback from anyone building agent systems.
The core problem we’re exploring is that AI agents are non-deterministic systems operating inside deterministic infrastructure.
Traditional systems assume actions run once.
Agents retry steps, re-plan tasks, and execute asynchronously.
That combination makes duplicate execution surprisingly easy.
Kybernis focuses on the execution boundary where agents trigger real mutations (payments, infrastructure changes, APIs).
Curious if others deploying agents have run into similar reliability issues.
What we kept running into is that even when the agent output itself is correct and admissible, distributed systems behavior can still produce duplicate mutations once execution starts — retries, worker restarts, async scheduling, etc.
So the layer we focus on is the execution boundary itself: once a tool call or API mutation is approved, ensuring that action commits exactly once.
Pre-execution validation definitely helps reduce bad decisions upstream. Our assumption is that even with good validation, the execution layer still needs deterministic guarantees because infrastructure failures are unavoidable.
Specifically, shared state management — agents reading and writing concurrently without collision detection leads to silent failures that look like model quality issues but are actually concurrency bugs.
We open-sourced a coordination layer for this: https://github.com/Jovancoding/Network-AI