OpenAI made computer use 10x faster in a year with a dual-agent Guardian setup
johncoogan · x · 2026-09-30
OpenAI's Thibault Sottiaux explains how the company achieved a 10x speedup in computer use over one year.
- Not just sampling faster: the gains come from both the harness and the model, especially reducing how much the agent must think before reliably taking its next action.
- Dual-agent architecture: a second agent called AutoReview (internally named Guardian) watches over the primary agent, and was also made faster.
- Harness engineering: the team used Astra at scale to improve harness performance, run more operations in parallel, and eliminate small stalls.
Combined, these changes produced the overall 10x speedup.
More from coding & agent
- Meta's Context Language Models treat context as an editable file, +65% on agent tasks — arankomatsuzaki · 2026-09-30
- Quarkslab builds agentic workflow that chains AI-found FreeRDP bugs into RCE — evilsocket · 2026-09-30
- Trust is not a security primitive: why AI agent sandboxes must assume nothing about the model — basedjensen · 2026-09-30
- Omni-Decision: evidence-ledger planning hits 81.4% on OmniGAIA at 43% of Gemini-3.1-Pro's cost — Ming Ma · 2026-09-30
- Omni-IO Skills: harness-level composition lifts GPT-5.6's omni-input support from 40% to 100% — NationalUniversityofSingapore · 2026-09-30
- TypeSafe's official Claude Code skill: batching 13 questions in one call is 12.2x cheaper and 10x faster — Roger_M_Taylor · 2026-09-30