Microsoft's ActiveSaddler picks training tasks by open failures: +7.5 on Terminal-Bench 2.0
rohanpaul_ai · x · 2026-10-05
A new Microsoft paper introduces ActiveSaddler, an active task-selection method for agent harness optimization. Instead of a fixed task list, it tracks failure patterns and targets the most valuable unfixed failures, or explores unseen tasks to discover new ones. Gains: +4.4 points pass rate on GAIA2 and +7.5 on Terminal-Bench 2.0; reaching 58.5% GAIA2 dev accuracy cost $298 vs $1,360 with fixed ordering. Takeaway: aim your optimizer budget at still-open failures.
More from coding & agent
- AI security workflow: repo scanner plus log-reading bot triaging vulnerability PRs daily — saranormous · 2026-10-24
- Having agents bring recommendations with items needing your input, 'I've reinvented Management' — josh_wills · 2026-10-24
- Agents made screwing around so cheap that 'finding out' stopped being exciting, dev says — zack_overflow · 2026-10-24
- Dev gives up on Google's Argon coding model, makes it write an apology handoff prompt — lowiqdoctor · 2026-10-24
- Building a patient-triage AI agent: handoff and compliance logs are harder than LLM reasoning — Frid-267 · 2026-10-24
- Codex is now fully usable inside the ChatGPT iOS app — dkundel · 2026-10-24