Skip to content

Kando

Kando logo

কাণ্ড — Bengali for event, incident, episode.

A production runtime for long-running agents where the event log is the agent, not a debugging artifact. Append-only log in, projected world out, reactive responders in between.

What is Kando?

Kando is an opinionated production runtime that wires two foundational projects together:

  • ActiveGraph (Yohei Nakajima) — the event-sourced reactive graph model for agents. Provides agent-native abstractions: projected world state, reactive responders, typed edges with semantic logic, fork-and-diff, causal lineage.
  • EventStoreDB — the event-native database. Provides the production substrate: native append-only streams, server-side projections, persistent delivery, cluster consensus, and a decade of operational hardening.

Neither is modified. Kando is the layer that wires the architecture to the infrastructure.

Key properties

Property What it means
Ledger-native Every agent run is an append-only event stream. No mutable state. No conversation context that outlives its events.
Deterministic world World state is always a deterministic projection of the ledger. Replay produces the same world, every time.
Reactive responders Functions subscribe to event patterns. When an event matches, the responder fires and emits new events. No orchestrator.
Cheap branches Fork any run at any position. The shared prefix costs nothing to replay. Only the divergent tail executes.
Causal traces Every event records what caused it. The full causal chain from goal to artifact falls out of the architecture for free.
Budget-safe Hard limits on events, LLM cost, wall-clock time, and recursion depth. Enforced by the runtime, not by responders.

Get started in 60 seconds

git clone https://github.com/ucalyptus/kando.git
cd kando
make start

Then run your first agent:

make setup
kando run kits/diligence --goal "Evaluate Stripe"
kando run kits/research  --goal "Understand quantum computing"

→ Full Quickstart

References