Devin Owen

Looks rad. When a specialist's reasoning chain breaks down mid-task, can you fix the broken step in place and have everything downstream re-evaluate, or does it effectively require a restart from that point?

Looks dope guys. What are cases you think this would outperform (either in accuracy, cost, latency, etc.) something like Claude Code subagents or agent teams?

The separate agent per role thing is a smart setup. Do they actually think differently (have substantially different context or use different models) or is it more about scoping what each one focuses on? How specialized does each one get under the hood?

Hiring fairly but quickly is a tough nut to crack, I'm glad that there are some new approaches. My previous startup was all about fair hiring so I really resonate with this problem and your solution. Biggest issue I've seen is how bias can creep in based on the underlying LLM, even if your system prompt says "don't be biased" haha. What are you doing to mitigate that?

Really like how you’re treating tests as something the agent owns end to end, instead of sprinkling “AI help” on top of existing tools. The interesting tension I keep seeing is between giving the agent freedom to refactor or regenerate tests vs preserving the team’s hard-won testing conventions and invariants. Curious how you handle that in practice once TestSprite starts touching a large,...

Cool to see you going all the way into “real actions” and not stopping at drafting replies like everyone else. The hard part I keep seeing with this kind of product is the tightrope between autonomy and blast radius: the more you let the agent actually change accounts or billing, the more one bad config can quietly hurt real customers. Curious how you’re drawing that line in v1? Especially...

Turning “PyTorch in, tuned CUDA/Triton out” into something productized like this is a very ambitious swing, especially with 32 agents coordinating on the same kernel. The hardest part of these systems in my experience is not just finding a faster variant once, but keeping the optimized kernels robust across driver changes, new GPUs and slightly different input shapes without a constant...

Really like that you’ve framed the interface as “manage a team from a doc” instead of yet another chat or IDE sidebar; that mental model feels closer to how senior engineers already reason about work. The big tension I keep seeing with agent tools is between hiding complexity so it feels magic vs exposing enough of the chain-of-thought that people can actually debug when things go sideways....

The synthetic replacement angle here is a clever way to meet security folks where they are without completely wrecking the day-to-day UX for people who just want to use AI tools. One pattern I keep seeing is security teams wanting strict redaction while ICs need rich enough context for the model to stay useful, and those two incentives often collide. How do you decide what level of semantic...