Agentic Harness with Jev: 3x Fewer Iterations, 444x Cheaper, Zero Hallucinated Files
Roger_M_Taylor · x · 2026-09-25
A thread + quoted builder's guide explains an agentic harness built with Jev: the lead model (claimed Opus 5.5) holds the whole codebase in context while eight executor agents make typed decisions — files, commands, and a 0–1 confidence score — instead of paragraphs. The model can only pick from allowed options, every decision is verified before touching code, and a memory layer preserves state across agents.
Results claimed: 8 agents close a task in 3 iterations instead of 10, 193x faster and 444x cheaper, with zero hallucinated files; output includes code, tests, docs, confidence and next steps. The quoted article is a step-by-step guide with full code, architecture and diagrams.
Related event: Developers Build Agentic Harness with Jev, Slashing Cost and Time(3 posts)→
More from coding & agent
- Demo: leave comments on a paper and let your agent address them — Rasmic · 2026-09-25
- Jev becomes fastest-adopted AI model ever, spawning sub-cent browser and trading agents in 3 days — multiply_matrix · 2026-09-25
- Making Games for Agents Means You Have Instant AI Playtesters — One Prompt Upgraded the Graphics — tristanbob · 2026-09-25
- DeepLedger Launches AI Staff Accountant MCP Connector for QuickBooks — modelcontextprotocol · 2026-09-25
- AgentWallet MCP Gives AI Agents Wallet Infrastructure Across Solana and EVM Chains — modelcontextprotocol · 2026-09-25
- New SDK Tests AI Agents on Real Websites and Workflows, Not Mocked Tools — Able-Ingenuity8885 · 2026-09-25