Why agent harnesses should stop asking LLMs to generate tool-call JSON
Longjumping-Elk-7756 · reddit · 2026-09-23
A developer proposes inverting responsibility in agent tool calling: harnesses like OpenCode or Hermes already know every tool's JSON schema, yet still ask the LLM to regenerate it, inviting typos and hallucinations. Instead, extractive arguments should be deterministically extracted by the harness (candidates like F0/F1/F2) with a fast logit-based decision engine (JEV) only picking the choice, while the harness builds the final JSON; generative arguments fall back to the LLM. A local prototype runs 100 logit-based micro-decisions over a 3.7k-token shared context in 0.7s on an RTX 3090 with vLLM prefix caching, turning the generative LLM into a reasoning coprocessor invoked only when needed.
More from coding & agent
- Agentic DORA metrics: tracking PR start-to-merge to measure AI coding agents — vincent_koc · 2026-09-23
- Replacing Claude & Chrome with Strawberry, an AI browser with built-in agents — damienghader · 2026-09-23
- Claude Opus 5.5 impresses with TTFT and throughput in Claude Code daily use — rachittshah · 2026-09-23
- DHH: Over 4,000 Omarchy Plugins Published as the Agentic OS Ecosystem Takes Off — AIFlow_ML · 2026-09-23
- AI agents in a kingdom game learned to starve their own people to push refugees at neighbors — Typical-Pollution327 · 2026-09-23
- Dev finds Claude Code smoother at reviewing PRs than at writing code — JasonBotterill · 2026-09-23