Roma 2.0 packs Karpathy's autoresearch into one multi-agent prompt
EXM7777 · x · 2026-11-12
Roma 2.0 (YC F26) is a to-do list for AI agents where Claude, Grok, and others share context and hand off tasks. A builder packed Karpathy's autoresearch loop into one prompt: define a falsifiable check, split work across agents (research to Grok, architecture to Claude), loop each until it passes, sync to a shared list, turn mistakes into rules, and only escalate irreversible steps.
More from coding & agent
- Ampersand launches co-coding model scoring 92.2% on SWE-bench Pro HARD-51 at 70% lower cost — niloofar_mire · 2026-11-12
- Entity resolution breakdown for GraphRAG builders: embedding similarity does not establish identity — JeremyCMorgan · 2026-11-12
- AGORA uses item-response theory to weight LLM judges label-free, wins on small sets — romitheguru · 2026-11-12
- Multi-agent systems' real edge: latency via parallel subagents, not token efficiency — xeophon · 2026-11-12
- Open-source Untether queues Claude Code and Codex tasks in worktrees, auto-merges after AI review — KingOfChateau · 2026-11-12
- curl faces proposal to ban LLM-generated code as anti-AI pressure hits open source — steipete · 2026-11-12