Linear vs Range

AI Research Roundup: Terminal Agents, Cloudflare Traffic, Nanochat

This newsletter edition covers recent AI research and tools. Key items include a paper on terminal agent training with self-improving tasks, a method for compressing agent screen memory, a tool for generating editable 3D scenes, and a model that predicts environment responses. Cloudflare's analysis of 206 million web sessions reveals mixed human-agent control, impacting bot detection. A new C file implementation runs a 744B parameter model efficiently, and a code graph tool supports 150+ languages. Additionally, Karpathy's nanochat project trains a GPT-2-class model for $48, and a benchmark shows Apple's SpeechAnalyzer outperforming Whisper Small. The newsletter also highlights videos on model serving and an internal agent at Linear.

• Cloudflare recorded 206 million Precursor evaluations in a day, showing sessions can shift between human and automated control.
• Colibri runs a 744B parameter MoE model from a single C file, using VRAM, RAM, and NVMe as tiers.
• Karpathy's nanochat trains a GPT-2-class model for about $48 on eight H100s.

Coding sessions: Linear Agent can now set up, run, and test your code

Linear Agent can now set up, run, and test your code before returning its work. That means fewer handoffs and changes that are further along when they come back to you.

AI Team Manifest Can Pass Validation But Be Unsafe

The article explains that JSON Schema structural validation can confirm a manifest's shape but cannot guarantee that an AI team configuration is safe or executable. NexFlow uses YAML manifests mapped to JSON Schemas and performs structural checks; the project also runs bounded semantic reference checks across manifests, but full semantic validation and runtime enforcement (permissions, approval gates, credential isolation, auditing) are distinct layers that a production runtime must provide. The author argues tools should report which validation layers they ran and which guarantees remain unverified to avoid misleading users with a single "valid" indicator.

• NexFlow describes AI developer teams through YAML manifests that map each supported manifest kind to a JSON Schema.
• At the reviewed repository checkpoint, `npm run validate` validates 113 manifests against 17 schemas.
• The repository includes a bounded semantic reference smoke check; `npm run semantic-smoke` passes for seven example projects at the same checkpoint.