Omni-IO Skills: open-source harness lifts agent multimodal support rates from under 40% to 100%
_akhaliq · x · 2026-10-01
A new open-source paper introduces Omni-IO Skills, a plug-and-play multimodal harness that wraps general-purpose agents like Claude and Codex so they handle audio, video, 3D and cross-modal workflows without retraining the base model. Key points:
- Hierarchical skills: 27 skills covering 38 tasks across 7 artifact modalities and 4 capability families (understanding, generation, reasoning, retrieval)
- Declarative execution graphs: multi-step workflows planned as dependency graphs; independent tasks run in parallel
- Asset registry: intermediate outputs persist and are reused across turns
- Pluggable backends: switching providers is a config change
On the UniM benchmark with GPT-5.6 Sol and Claude Sonnet 5 as hosts, support rates rise from under 40% to 100% and SQCS from 27% to 75–78%. Code and 20 example workflows are open-sourced.
More from coding & agent
- Tweet 'I need an app that...' and someone ships you an MVP in 30 minutes — thejasminejade · 2026-10-01
- 7-person robot startup Innate says Grok and Cursor let them do the work of 50 — Baconbrix · 2026-10-01
- Dev uses OpenAI's Dot to seamlessly monitor Claude Code benchmark on his machine — dkundel · 2026-10-01
- ChatGPT Sites Can Now Host MCP Servers, Auto-Deploying Plugins Across Platforms — OpenAIDevs · 2026-10-01
- Omnigent 0.16 ships with new file browser, admin controls and smoother onboarding — matei_zaharia · 2026-10-01
- Rox benchmarks: Jev reranking beats GPT-5 Mini — 20x faster, 10x cheaper, 12% more accurate — hardimanjames · 2026-10-01