- Author
- Manish (InstaVM blog)
- Published
- 10 November 2025
- Kind
- Company blog post of personal practitioner anecdotes
- Original
- instavm.io/blog/llm-anti-patterns
- Subject
- Harnesses and agent loops
What it is
Manish writes on the InstaVM company blog. The post shares lessons from 15 months of working with LLMs. It defines anti-patterns as behaviors to avoid.
What it says
- Agents should not resend near-identical context, such as screenshots that barely differ.
- Ask the model to write code for exact tasks, such as counting Rs in BLUEBERRY.
- An image-text task failed after about 13 tries but worked via a Google Doc.
- Long sessions suffer; in one the model forgot a given database URL and supplied another.
- Models do worse on obscure topics and on topics newer than their training.
- Vibe coding loses the thread; the author caught a password_hash exposed in an API response.
What evidence it gives
The evidence is the author's own anecdotes, with screenshots of some cases. It links a few outside pieces, including one on context management. It gives no measurements or repeated tests.
Interests and limits
InstaVM sells code execution and computer-use infrastructure, and the post promotes its own open-source tool. Each lesson rests on one or two incidents. The screenshots were not viewed. The claims about context size and accuracy are the author's impression.