You've nailed the real friction point that demos gloss over: agents are great at generation but terrible at verification in production systems. The vision latency tax is brutal once you hit real workflows.
This addresses a real pain point—runtime guarantees vs probabilistic hopes. A few questions from someone who's dealt with LLM guardrails in production:
1. How does CSL handle the gap between what an LLM intends to do (based on its reasoning) and what constraints allow? For example, if a policy forbids "database modifications" but an agent legitimately needs to write logs—does the DSL let you express intent-aware exceptions, or do you end up with overly broad rules?
2. Z3 constraint solving can be slow at scale. What's your performance profile when policies are deeply nested or involve many symbolic variables? Have you profiled latency on, say, 100+ concurrent agent requests?
The formal verification angle is solid, but I'd be curious whether you've stress-tested the actual bottleneck: not the policy logic itself, but the interaction between agent reasoning and constraint checking when policies need to be permissive enough to be useful.
This is a clever use case for LLMs, but I'd be curious about the quality of signal you're extracting from Discord noise. How are you handling the challenge of distinguishing genuine feature requests from off-topic conversation or jokes? Are you doing any filtering/validation before feeding to the model, or relying on the LLM to handle that context-switching itself?