- Author
- Posted by mr. TIM (timkellogg.me), with replies from several users
- Published
- 6 August 2026
- Kind
- Bluesky post with a short reply thread
- Original
- bsky.app/profile/timkellogg.me/post/3msftxqpmfk2q
- Subject
- Harnesses and agent loops
What it is
The author posted a few lines about Prime Agent and recursive language models. Six replies followed, and the author answered several. The page covers the post and the replies that were returned.
What it says
- The post thinks Prime Agent is likely a watershed, as a better harness captures overhang.
- It says no one has yet built a coding agent on recursive language models.
- It says it seems likely that RLM has replaced the ReACT pattern for big problems.
- The author's reply says the agent gets the prompt as a Python variable.
- The author's reply says compaction cleans cell outputs but leaves all variables intact.
What evidence it gives
The post gives opinion and no data. The author's replies add a few mechanism details. One user reports not seeing the described behavior in testing, and the author disputes this.
Interests and limits
The source is short, and this page does not guess beyond it. It is one person's view of one tool. Prime Agent is never linked or defined in the text reached. A quoted post in one reply was not viewed.