AI agents are writing their own instructions, and the clutter may be making them worse

AI agents that rewrite their own instruction and memory files can quietly degrade their performance, according to Jamie Watters. The practical fix is decidedly unglamorous: keep a person in the editing loop, review the files regularly and delete the bits that no longer earn their keep.

Watch Desk analysis

What happened

In a report published on 21 September, Watters says agents that autonomously update their own guidance accumulate stale instructions, guesses recorded as facts and unnecessary clutter. The article attributes benchmark findings from SkillsBench to the claim that model-written skill files can perform worse than having no instruction files at all, while human-curated files improve task success rates.

Watters recommends a monthly manual review. The proposed housekeeping is simple: retain useful rules, remove noise and unconfirmed guesses, and date instructions so people can tell how fresh they are. That gives operators something concrete to do instead of merely hoping an agent’s digital filing cabinet develops standards.

Why it matters

Instruction and memory files increasingly sit between an agent and the work it is meant to complete. If those files become a scrapbook of old assumptions, the system may follow yesterday’s workaround with complete confidence today.

The finding also challenges a popular instinct in agent design: that more accumulated context must be better. According to the account, carefully selected guidance can help, but automatically generated guidance may add friction or actively reduce success. For teams deploying coding, research or operational agents, the quality of the instruction file deserves testing alongside model choice and tool access.

Our read

Treat self-updating agent notes as a draft, not as institutional memory. A monthly review is a sensible starting rule, but teams should also keep a record of what changed and test whether a new instruction improves the task it claims to improve.

The supplied evidence is Watters’ account of the SkillsBench findings, not the benchmark’s full methodology or an independent replication. That matters. Still, the operational lesson is strong even before the grand theory of agent memory is settled: an instruction file should be edited like software, not worshipped like an archive.

What to watch

• Whether SkillsBench publishes the underlying tasks, participant models and full comparison between model-written and human-curated files.
• Whether the performance penalty comes from stale information, sheer file length or incorrect instructions.
• Whether agents can propose changes safely while leaving approval and deletion to people.
• Whether teams begin measuring instruction-file quality as part of routine agent evaluations.

Discussion spark: Should AI agents ever be allowed to edit their own instructions without approval, or is a human review only useful once the damage has already been done?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.