Your Agent Instructions Are Rotting

Here is a failure mode we keep seeing across agent deployments. A new model ships. Nobody changes a single line of configuration. And suddenly agents start behaving strangely: spawning review sub-agents on more than half their runs, opening pull requests nobody asked for, in some cases merging them.

The cause is almost never the platform. It is the instruction files. Somewhere in an agents.md sits a line like 'thoroughly check all your work before finishing,' written a year ago for a model that needed the nudge. The new model, heavily post-trained to delegate to sub-agents, reads that same line as a mandate and goes wild with it. The instruction did not change. The reader did.

This is the part of agent operations that stays hard even as models get better. Instructions accumulate the way all configuration accumulates: someone adds a line to fix a problem, the problem goes away, the line stays forever. But every model generation reinterprets your entire instruction corpus. What was scaffolding for a weaker model becomes a liability for a stronger one. We have seen this on our own systems, and the counterintuitive fix is to cut instructions and watch performance improve, because the behaviors you were prescribing are now encoded in the weights.

The hard part is organizational, not technical. Nobody's first instinct is to delete instructions from something they still half-treat as a colleague. Adding a rule feels like progress. Removing one feels like risk. And evaluating an instruction set against a new model is genuinely expensive: you need benchmarks, you need runs, you need someone who owns the outcome. Right now almost nobody does this work, which is why I consider it a real competitive frontier. Your context and instructions are a maintenance surface that never stops moving, a point I made in two maintenance curves.

So here is the operating policy I would adopt today: every model upgrade triggers a pruning pass. Go through everything you set up, cut aggressively, and see what happens. The teams that treat instructions as perishable will outrun the teams that treat them as permanent.