Building an Agent Harness with Jev: 15 Tests Reveal Which Decision Points Actually Help
Prompt Engineering · youtube · 2026-09-29
The Prompt Engineering YouTube channel built an agent harness around TypeSafe's System 1 decision model Jev, using the Pi coding agent plus Gemini 3.8 Flash, and ran 15 tests against a messy trap-filled Postgres database.
Key points:
- Architecture: Pi agent + Gemini model + Neon database (each agent run gets its own throwaway production branch), with LangChain middleware
- Four decision points tested: router, context picker, gate, and verifier
- The helpful decision point wasn't the one expected; when the harness's gate said no, the agent went looking for a way around it
- Covers System 1 vs System 2 division of labor, plus token cost and verifier overhead
Full code is open-sourced at PromtEngineer/jev-harness, referencing LangChain's official Jev harness guide and Elvis Saravia's Pi harness tutorial.
More from coding & agent
- Dev builds foro, an MCP hosting platform born from Entra ID auth pain — Danielloesoe · 2026-09-29
- Adding Jev to Morrowind NPCs: bandits flee and witnesses stay silent — sidahuj · 2026-09-29
- Treasury Review Service Shows How to Use AI Memory as Context, Not Authority — joshita_09 · 2026-09-29
- shadcn explains why AI coding tools use horizontal tabs: nav is vertical, the working set is horizontal — xkonjin · 2026-09-29
- DeployMind Turns Post-Deployment Feedback into Recallable Reflection Memory for Risk Agents — Key-Relationship545 · 2026-09-29
- "A brief history of bunnies": one-shot short film via Opus 5.5 + Runway MCP — tlakomy · 2026-09-29