New Paper Measures Time Awareness in LLM Agents, Astra Leads
maksym_andr · x · 2026-10-08
A new paper systematically measures time awareness in LLM agents, covering duration following, forecasting, and retrospection.
- The authors find Astra is strikingly better at duration following than any other model (Figure 1), likely tied to training on strictly time-bounded tasks, including those behind recent agentic incidents; the higher time awareness may be an instrumental skill learned almost by accident through task time constraints.
- The same does not hold for Anthropic models: Fable 5.1 does not reliably follow any given duration—asked to work for 4 hours, it usually finishes in under 1 hour.
- The authors argue time awareness has long been overlooked but matters greatly, especially for long-running agents.
Related event: AgentTime Benchmark Measures LLM Agents' Time Perception; Astra Leads(2 posts)→
More from coding & agent
- Forbes debuts Agentic 20 list spotlighting AI agents doing real-world work — annatonger · 2026-10-09
- Agent Plasticity: top-performing agents aren't the most efficient learners — RulinShao · 2026-10-09
- Job Search OS in Claude Code: 20 minutes a day beats 3-hour mass applications — aakashgupta · 2026-10-09
- Solo Founder Builds Entire Business With 7 AI Agents Working in Parallel — PrajwalTomar_ · 2026-10-09
- img2threejs: open-source image-to-3D rebuilds reference images as procedural Three.js models — tom_doerr · 2026-10-09
- Dev builds interactive teardown of a humanoid robot with 56 actuated joints using Claude Opus — techartist_ · 2026-10-09