On posts like these I would like to see the trade offs and why choices were made.
For instance, why create your own agent framework vs the 100s that exist including Google ADK. On logging, did you build your own telemetry system or did you use Otel and something like Langfuse.
If you built your own for all this, what were the reasons why?
Job security is an acceptable reason
Hey thanks for the thoughts! I can try to detail this a bit in a future post, I briefly touched on this in the intro but the reality is when we started this effort there just _werent_ many options in Java. Its possible this would weight differently today. But a lot of the answer is simply that building our own has allowed us super deep integration with existing (custom) infrastructure.
On telemetry, we do a combo. The transcript storage format we use in aviator itself forms a great tracing system so we often use this as the source of truth, but we also produce Otel spans for specific requests/responses and use tempo etc for querying/viewing that data.
Great post! Super cool for the future of harnesses.
On posts like these I would like to see the trade offs and why choices were made. For instance, why create your own agent framework vs the 100s that exist including Google ADK. On logging, did you build your own telemetry system or did you use Otel and something like Langfuse. If you built your own for all this, what were the reasons why? Job security is an acceptable reason
Hey thanks for the thoughts! I can try to detail this a bit in a future post, I briefly touched on this in the intro but the reality is when we started this effort there just _werent_ many options in Java. Its possible this would weight differently today. But a lot of the answer is simply that building our own has allowed us super deep integration with existing (custom) infrastructure.
On telemetry, we do a combo. The transcript storage format we use in aviator itself forms a great tracing system so we often use this as the source of truth, but we also produce Otel spans for specific requests/responses and use tempo etc for querying/viewing that data.