Minimal web frameworks (e.g., Express, Flask, ASP.NET Minimal API) are significantly more token-efficient for AI agents than full-featured frameworks, with a 2.9x token gap.
All 19 tested frameworks produced working blog apps without bugs on the initial build, highlighting rapid progress in agentic development.
Full-featured frameworks like SvelteKit and Django were the most efficient among their category, while Phoenix was the most token-expensive.
Adding features (e.g., categories) cost roughly the same token band (15-30k) across most frameworks, regardless of their initial build cost.
Framework overhead primarily impacts initial builds; extending existing code has similar costs everywhere.
ASP.NET Minimal API stood out for token efficiency, static typing, speed, and low memory use.
Popular frameworks like Rails and Laravel caused confusion for agents due to missing system packages.
The user proposes a tag-based EDIT tool for local LLM agents, using line-number and checksum tags instead of verbatim text to save tokens and increase reliability.
Current CAS (Check and Set) edit methods require the old text verbatim, causing token waste and failure with hallucinations or formatting issues.
The new approach returns lines prefixed with a 4-character tag (checksum) and allows edits by line:tag pairs, reducing token usage especially for deletions.
The tool is effective with models like DeepSeek v4 Flash, and the line:tag format helps LLMs exploit line information for ranges and follow-up calls.
An alternative is using a file-level CRC32 tag with only line numbers, which saves tokens but fails on any unrelated changes, making tradeoffs apparent.
The author suggests practical testing with ds4-agent across sessions to decide which method is better, with a command-line switch as an initial step.