DeepSeek open-sources Super-Agent, a proactive Claude Code rival
An MIT-licensed prototype that watches your repo and chat to open PRs unasked — plus OpenAI's SDK adds context compaction and labs align on a crawler opt-out.
Copy markdown
A coding agent that doesn't wait to be asked
DeepSeek's Harness team open-sourced Super-Agent (MIT license), a prototype that watches a GitHub repo and Feishu chat, spots coding problems, and opens a PR in its own session — acting only when asked or when a team rule permits. It's aimed squarely at Claude Code and Codex, but DeepSeek stresses it's a research prototype, not an official product, so audit the code before running it.
OpenAI's Python SDK learns to compact agent context
openai-python 3.42.0 lets you tune agent context compaction client-side instead of leaning on server defaults, and adds session review, tool settings, experimental time awareness, and org-level external key management. No model or price change — it's aimed at keeping long agent runs from overflowing the context window.
A robots.txt opt-out for AI training firms up
Major providers are converging on a voluntary robots.txt token (usage_ai-training) to signal per-path training preferences. OpenAI says it'll ship support "in the coming months"; Microsoft's preferences take effect Feb 1, 2027, and it stops reading NoArchive/NoCache as training signals after Mar 1, 2027 — so if your site relies on those tags, plan new settings before then.