LiteCode Exposes Small-Model Agent Failures

Hot take: small models are less useful as “smart chatbots” than they are as brittle components inside a heavily constrained agent pipeline, so most of the reliability work belongs in the wrapper, not the prompt.

• –The strongest claim is that markdown fences and JSON validity are operational problems, not prompt-engineering problems; post-processing and validation are mandatory.
• –The write-up is most convincing where it describes orchestration failures: wrong-file edits, function-name drift, and read-only requests being misrouted into destructive actions.
• –The practical takeaways are solid: pre-call token accounting, one-file-at-a-time editing, and short state summaries are all easy wins for agent reliability.
• –The article is less a benchmark and more an implementation guide based on repeated failures, which makes it useful for builders even if the conclusions are partly anecdotal.