Harness is dead? AI Engineer NYC debates stripping scaffolds vs. multi-day agents
TheTuringPost · x · 2026-10-19
At AI Engineer NYC the author heard sharp disagreement about agent harnesses: some engineers are stripping harness components away because models no longer need them, while others build systems letting agents work autonomously for days.
The resulting guide starts from a 25-line coding agent and draws on insights from philschmid, dillonrolnick, aparnadhinak, akshatb and a talk by jediahkatz, addressing: what an agent actually does during those days, how it remembers events without cementing early guesses as facts, and what we're really allowing it to change when it builds its own tools.
More from coding & agent
- Dev builds cross-platform hero shooter in under a week using Claude and Codex agents in background — garrytan · 2026-10-19
- AgentBase MCP Server opens a searchable registry of agent prompts, skills and workflows under CC0 — modelcontextprotocol · 2026-10-19
- Sovereign AI Blog ships an MCP server for its NVIDIA DGX Spark self-hosting log — modelcontextprotocol · 2026-10-19
- Berkeley-led paper: making LLM agents 'forget' history yields up to 39% gains in discovery tasks — burny_tech · 2026-10-19
- MinniMemoryMCP cuts Claude Code's per-turn memory load 43% by reorganizing context — MinniLabs · 2026-10-19
- Why agent sandboxes became a thing: slow VMs pushed Lovable, E2B and Modal to build their own — vboykis · 2026-10-19