AI Agent Apps

[gangtao]

1. using timeplus stream for all user input, llm input/response tool input/output
2. all tools are just UDF
3. all LLM calls using UDF as well
4. Agent Loop is just a running Timeplus stream with MV
5. memory as vector search

[Keesan12]

This architecture is interesting because it turns the agent loop into a dataflow problem instead of a framework problem.

The thing I would watch closely is whether the stream model preserves decision boundaries, not just events. In practice the expensive failures usually happen when the system has lots of telemetry but still cannot answer:

• what state is this run in
• what changed since the last attempt
• why is another step justified right now
• what condition ends the run

If every tool call and model response lands in streams, I would still want a small canonical run-state machine on top of that. Otherwise it becomes easy to replay history and hard to govern live behavior.

That has been a big lesson for us with MartinLoop-style runs: observability is necessary, but the real trust comes from making the loop prove why it deserves another turn.