Coding-agent prompt files grow 226% and never shrink

A preprint dated August 11, 2026 tracked 247,694 instruction lifetimes across 1,867 repositories and found that agentic coding instruction files grow without bound, more than tripling over their lifetime at plus 226%. The paper on Arxiv names the effect catastrophic remembering.

At a glance

• The asymmetry is economic: appending an instruction is always cheap, while deleting one whose rationale has been lost, without risking a correctness regression, costs O(2^|D|) in a prompt of |D| instructions.
• Files gain 4.9 net instructions per commit, and the older an instruction is the less likely it becomes to be deleted, with log-hazard falling 0.032 with each commit.
• Comments inside the prompt halt the growth: on inverted IFEval they remove 99.3% of excess instructions, cutting bloat from plus 211.3% to plus 1.4% against known optimal prompts.

Why it matters For teams running a coding agent against a long-lived repository, the finding reframes prompt bloat as a structural outcome rather than sloppiness. If the deletion cost really is exponential in the number of instructions, no amount of discipline fixes it, and periodic wholesale rewrites become the only lever left. The proposed alternative, treating a prompt file like source code with comments, appears cheap enough to test immediately.

The paper frames catastrophic remembering as the inverse of catastrophic forgetting, the failure mode around which continual learning is organized. In the repositories studied, growth halts in only two situations: the project retires, or someone rewrites the instruction file wholesale rather than editing it line by line.

To test a fix, the authors invert IFEval to build verifiable worlds whose optimal prompts are known in advance, which makes excess instructions measurable rather than a matter of judgment. In those worlds, comments that encode the latent reasoning behind each instruction bring prompt length back near the optimum.

Applying the same inversion to WildIFEval, the authors report that prompt comments improve real-world agentic instruction-following by up to 23.1%. The published abstract gives that figure as an upper bound and does not break it down by task, model or instruction count.

The submission is catalogued as arXiv:2608.11095, filed under artificial intelligence with cross-lists to machine learning and software engineering, and listed from Kushal Chakrabarti; version 1 runs to 110 KB. The abstract closes on the analogy that drives the argument:

If English is the new code, why don't we have comments yet?

What's next Whether comment conventions get adopted across agent harnesses remains open, and the arXiv record lists no venue for the work beyond the preprint itself. Also untested is whether the WildIFEval improvement holds outside the inverted setting, and whether annotating rationale at write time is cheap enough that maintainers actually do it.