We're building a CLOB-based DEX and this problem hits close, we've interviewed a lot of smart contract developers where AI makes it nearly impossible to assess real understanding. What we've shifted to: giving candidates a live matching engine with a subtle bug in the order execution logic and asking them to find and fix it. Prompting Claude gives you something that looks right but breaks under edge cases, partial fills, price-time priority violations, self-trade prevention. The candidates who actually understand the mechanics catch it. The ones steering AI without real knowledge submit a fix that breaks three other things. Your "AI steering" framing is exactly right. The real skill now is knowing when the output is plausible but wrong and that only comes from genuine domain knowledge.
The fragmentation problem across agents is real, we ran into the exact same issue managing rules across different dev environments. The unified view approach makes sense, but curious about sync conflicts: if you update a skill file via Cursor and then via Claude Code, does Skills Manager detect the diff or does last-write win? Also wondering if you plan to support team-level skill sharing, not just individual installs from GitHub repos.
The document standardization problem you're describing maps closely to what we see in DeFi infrastructure, different chains, different data formats, no consistent standards, and existing tools breaking on real-world inputs. The "model agnostic base + market-specific fine-tuning" architecture is smart. Curious how you handle cases where the same lender operates across multiple markets with conflicting document conventions, does the model layer stay separate per market or do you find cross-market signal bleeding actually helps fraud detection?
Communication infrastructure breaks first. When you're 10 people, context lives in people's heads and Slack threads. At 50, that stops working entirely. The fix isn't more meetings, it's written decision logs. Every significant decision needs a one-paragraph record: what was decided, why, and who owns it. Took us longer than it should have to figure that out.
reply