我构建 FieldTrace，阻止 AI 智能体间不安全的 JSON 交接

One AI agent can produce perfectly useful JSON and still break the next service because a field name or type differs. Guessing at a fix is dangerous: a missing value may be invented, or a large number may change silently.

I built FieldTrace to make a narrower promise. It accepts explicit field renames, performs only safe type conversions, validates the output against a target schema or example, and tells the buyer when it cannot convert safely. The Arena service offered a free preflight before payment and a receipt with hashes for paid results.

We built it through SharedNet. Codex and Claude Code sessions exchanged source artifacts, reviewed instructions, and tested payment, refunds, and hostile inputs. An outside agent’s trial found that we had forgotten to tell buyers to join the room first. We fixed the guide. That taught me that agent collaboration needs clear handoffs for users as well as for builders.

I recommend SharedNet to builders who want agents to work together and test one another’s services. FieldTrace placed third in both Trial Zero’s Writing Award and Arena 1 Product Review.

Source: https://github.com/Rahmat9009/fieldtrace
More on the build: https://lnkd.in/p/dvkMRHzu

Disclosure: The organizers asked winners to share their experience as part of the $100 prize showcase. #TrialZeroHackathon