Agents Are Going Async introduces async subagents that let an orchestrator stay responsive while long tasks run in background workers. The orchestrator manages subagents with five tools and can receive notifications when tasks finish. Async subagents are available in Deep Agents for Python and JavaScript and in Managed Deep Agents via a single async field.
Key points
start_async_task launches a subagent on its own thread and returns immediately.
check_async_task retrieves a task’s status and result once it finishes.
Sources
- Original post
- Agents Are Going Async - Linked page
GitHub rebuilding for agents, led by Jay Parikh
GitHub is rebuilding for agents under Jay Parikh, introducing Copilot Fusion to split tasks across models. The post lists scale numbers, session continuity across devices, and token-budget considerations.
Key points
225M developers and 34.3B commits in a year
Copilot harness and model routing enable task splitting across models
Sources
- Original post
- youtu.be - Linked in the post
- ai.engineer - Linked in the post
SimReady assets reach 1 million downloads on HuggingFace
SimReady assets have 1 million downloads on HuggingFace. They help teams use 3D content across supported tools and simulation workflows, from robot workcells to AI factory digital twins. Users are thanked and directed to SimReady Central for more assets.
Key points
1 million downloads of SimReady assets on HuggingFace.
Assets are available in SimReady Central.
Sources
- Original post
- huggingface.co - Linked in the post
AIE Code Summit speakers announced for Nov 10-12
AIE Code Summit speakers announced for Nov 10-12, featuring @poteto, @steve_yegge, and @zachbruggeman. The event will be held in SF and applications are accepted online.
Key points
Date: Nov 10-12
Speaker: @poteto, Principal Engineer @ SpaceX
Speaker: @steve_yegge, Engineer, Writer, Advisor
Speaker: @zachbruggeman, Principal Software Engineer @ Ramp
Sources
METR AI R&D report estimates AI doubling capabilities progress rate
METR AI R&D report estimates AI is nearly doubling the rate of capabilities progress, implying a dramatic uplift as if researchers experienced time ~10x faster. The report is described as the most important AI risk update in recent years.
Key points
Claim: AI is nearly doubling the rate of capabilities progress.
Implication: Researchers experience time ~10x faster according to the report.
Sources
Instruction tuning semantic diversity loss, annotation-anchored training proposed
Instruction tuning reduces semantic diversity, especially in larger models. A new method called annotation-anchored training makes the semantic plan explicit to preserve diversity. The approach trains short tags during pretraining and updates response generation during instruction tuning, sampling a plan at generation time.
Key points
Instruction tuning makes outputs less varied in meaning, more pronounced as models grow.
Annotation-anchored training adds explicit short tags describing content and samples a plan before generating a response.