AgentTime Paper: Agents Learn to Control Their Own Runtime and Forecast Task Durations
maksym_andr · x · 2026-10-09
maksymandr points to a detailed thread on AgentTime, a new paper teaching agents time-related skills:
- Self-controlled runtime: agents decide when to stop working
- Duration forecasting: models predict how long a task will take
- Time perception: agents understand how much time has passed
On duration-following, GPT-6 Astra beats Fable 5.1, 63% to 4%. Amusingly, in some runs Astra prefers to sleep rather than continue working.
Related event: AgentTime Benchmark Measures Temporal Perception in LLM Agents(3 posts)→
More from coding & agent
- Voyager launches as an open harness for AI-driven video, graphics and games — jgooten · 2026-10-09
- TermGrade: 1k open-source executable terminal-agent RL environments with full training recipe — maximelabonne · 2026-10-09
- Step 5 Preview free in Cline for a week, beats Kimi K3 and GLM-5.3 on DeepSWE — StepFun_ai · 2026-10-09
- Open-source idea-to-launch prompt turns vague ideas into full go-to-market plans — Andrew0_0 · 2026-10-09
- Thrixel launches AI 3D creation engine that lets coding agents build editable, interactive 3D worlds — RanaHanocka · 2026-10-09
- Where to run your agents: Mac for quick start, a server for the long run — lucasmeijer · 2026-10-09