MIT open-sources VISTA visual harness, enabling Claude Opus 5.0 to solve all 25 ARC-AGI-3 games
JFPuget · x · 2026-09-06
MIT researchers (with Kaiming He as co-author) have open-sourced VISTA, a visual harness that gives general-purpose multimodal models a continuous visual interface for long-horizon reasoning in interactive environments.
- How it works: every environment frame is archived as visual memory, letting the agent revisit original evidence while reasoning and acting in an observe → reason → act → observe loop.
- Results: paired with Claude Opus 5.0 (via Claude Code at xhigh effort), VISTA completes all 25 public ARC-AGI-3 games with a 100% win rate and a perfect Relative Human Action Efficiency (RHAE) score of 100. Codex CLI with GPT-5.6 Sol reaches 99.
- Code is available on GitHub, with an accompanying blog post.
More from coding & agent
- Hermes Agent adds per-model provider pinning for OpenRouter users — Teknium · 2026-09-06
- Open-source MCP server lets Claude Code, Codex and Cursor search each other's chat transcripts — Stormix4 · 2026-09-06
- PaperCompiler compiles papers into file-level specs to fix lossy paper-to-code agents — mohitban47 · 2026-09-06
- Designing Agent Context by Scope and Time: A Practical Guide for Agent Builders — bibryam · 2026-09-06
- Dev take: multi-agent setups don't help most tasks with current models — BLUECOW009 · 2026-09-06
- Four Prompts, Five Minutes: Codex Composes a Concerto with Scores, Animation, and Real Instruments — mhmazur · 2026-09-06