Axioms

The design of ikidna is informed by a few axioms chosen to guide the architectural principles needed for the system.

1. Recoverability and human support

Humans should always have an escape hatch so they can jump in and recover something manually if need be. While the preference is that everything is automated, we always need to make sure that a human can also administer it and recover the system if something catastrophic happens.

2. Flexibility

AI development is incredibly fast, and we’ve seen huge improvements over the last couple years. Whatever architecture we make, cannot commit to being overly opinionated, lest new research introduces improvements that we cannot adapt to. Similarly, one cannot lock themselves into a single model provider, and we need to have redundancy in the case that model inference providers are not available. That cannot take the entire system down.

3. Telemetry and observability is key

Any sufficiently large system needs significant amounts of telemetry data in order to isolate when problems occur. For an autonomous system that is expected to look after itself, it also needs The ability to access its own telemetry, both to improve itself and also to address any issues that arise.

4. Long-term optimization

This agent handbook espouses that we should be using frontier models first when we are encountering situations and scenarios where there is uncertainty and no pre-existing pattern. We don’t pretend to ignore the fact that constantly running expensive frontier models would be incredibly expensive. However, much like in human research and interaction We suggest using skilled models initially and then using them to lay the groundwork with implementation artifacts, tools, best practices, and deterministic workflows such that we don’t solve the same problem over and over and are able to progressively move tasks down the intelligence ladder. See the Cynefin framework for a similar conceptual model.