I would guess that it's closer to good old-fashioned plagiarism: someone gave the LLM your post and said "rewrite this in my voice" or something.

Seems way more likely than an LLM is regurgitating close-to-verbatim content from its training set.

Looks amazing!

Number one "complaint" I keep hearing about gridfinity - once you start you really want to do All The Things, and the print times are bruuuuuttal. Someone should sell injection molded snap-fit grids.

sub-agents in particular are a big deal, because each agent uses it's own temporary context window, keeping the main context window clean.

Claude Code also uses this trick for it's built in tools (e.g. WebFetch)

lots more detail in the full post: blog.thepete.net/blog/2025/12...

each coding agent can bring:
- different system prompt
- different tools
- (hidden!) context management tricks to steer the model’s behavior
- task management tools to nudge the LLM into better planning
- sub-agents, which can have a *huge* impact on context management

🧵...

turns out that coding agents are much more than a “wrapper” around an LLM. They do a surprising amount of work under the surface to (a) maximize the useful information available to the LLM, and (b) steer its behavior.

🧵...

turns out that:
- your agent can burn through it fast
- your LLM's performance starts taking a hit way before that "limit" is reached

further reading:
- blog.thepete.net/blog/2025/10...
- research.trychroma.com/context-rot
- github.com/adobe-resear...
- arxiv.org/abs/2404.06654

I worry a lot that this will induce more `LGTM` reviews from human reviewers when they see the "comprehensive" feedback from an LLM.

Particularly when coupled with the increased overall volume of PRs that AI-assisted coding is producing... 😬

The biggest risk with AI code review is that it looks so legit! Each review is so thorough and detailed.

No surprise: The LLM does a *great* impression of the huge pile of human code review examples in it's training set!

This doesn't mean that you shouldn't use AI code review. But you should treat it as an additional piece of feedback on a PR, rather than any sort of replacement for human review.

Think of it as a super-powerful static analysis step in your CD pipeline.

Also, given that
- the LLM won't have the full context of the codebase, and
- an LLM only has mediocre software design skills (at best)

you're not going to get any sort of full critique of your proposed solution from an AI-powered tool.

See if you can spot the prompt injection: github.com/trailofbits/...

Or if you can spot the backdoor being added: github.com/trailofbits/...

More on the lethal trifecta: simonwillison.net/2025/Jun/16/...

c) invest in making your codebases agent-friendly by wrapping all your conventions and tribal knowledge into a small set of scripts

point C is exactly the same guidance that I was giving 11 years ago(!!!): www.thoughtworks.com/insights/blo...

I was ahead of the curve I guess :D