AI Daily Digest · 2026-11-21
🔬 New AI Craft
1. Challenge regression diagnoses independently and verify recovery after deployment
Performance Factory persists task state across investigation, proposed fixes and post-deployment checks, waking agents on review or deployment events. A separate session challenges the diagnosis; an engineer decides what merges before an agent verifies recovery against baseline. In a manual sample of 100 root-cause suggestions, 70% were graded accurate, underscoring the need to establish causality rather than count investigations.
https://openai.com/index/how-openai-performance-factory-keeps-chatgpt-fast
2. Train LoRAs inside deployed agent harnesses and validate the learning signal
Hugging Face trains Qwen3.8-27B LoRAs inside the actual Pi, OpenCode and Hermes loops, checking tasks with reference and wrong-answer controls and filtering for partial baseline success. Coding evaluations within the target harness show accuracy and efficiency gains. AutomationBench gives no clear transfer to a different domain and loop, so the savings should not be generalized to arbitrary workloads.
https://huggingface.co/blog/harness-adapters-at-scale
🛠️ Tools & Tips
1. Reuse Codex host sandboxing for standalone commands and explicitly wrap MCP servers
OS-level enforcement constrains command process trees while keeping the existing host toolchain available; codex sandbox exposes the same permission profiles to standalone commands. Stdio MCP server launch commands are not automatically sandboxed this way. Constrain them explicitly when needed and validate the profile rather than assuming all tools share the command sandbox.
https://openai.com/index/inside-the-codex-sandbox
⭐ Open Source Highlights
1. Inkling 2 Mini opens Apache-2.0 multimodal weights with a looped core
Inkling 2 Mini takes images and text and produces text, reusing its middle 12 layers three times for 48 layer visits from 24 layers of weights. The model card specifies Apache 2.0 and up to 1M context. Looping adds per-token compute, and minimum reasoning effort does not fully disable reasoning.
https://huggingface.co/thinkingmachines/Inkling-2-Mini
📰 Industry News
(no items this day)
🚀 Major Releases
(no items this day)

