Technical Context
I got drawn into this discussion not because of the fancy phrasing, but because it directly impacts the cost and quality of AI automation. When an agent drags all the historical junk along, it doesn't think better. It just makes more expensive mistakes.
The idea itself is simple: in compression, I first keep the invariants, then extract the high-information-density essence, and for complex parts, I show an example instead of a long explanation. That is, I don't retell the entire plan but preserve what must not be broken, what affects the decision right now, and what a good result should look like.
And this is where I understand both sides of the argument. If the plan is flawed from the start, you can indeed end up with hundreds of lines of discarded code. But a detailed review of every plan can also easily turn into a token sinkhole, where the agent spends context on introspection instead of working.
In practice, I would separate two layers. The first, permanent layer: goals, constraints, architectural prohibitions, critical assumptions. These are the invariants. The second, short-lived layer: the current step, controversial decisions, fresh signals from logs, failures that must not be repeated.
I interpret the part about 'high perplexity' from an engineering standpoint, without the romance. You should keep not the 'smartest' thing, but the rarest and most useful: an unexpected bug, a hidden API limitation, a conflict in requirements, the cost of an error. The agent will generate all the mundane stuff on its own. It will forget everything unusual first.
And 'Show, don't tell' works great in prompts. Instead of saying 'write briefly and to the point,' I'd rather give a mini-example of good compression. The model picks up the format faster, and I get less stylistic drift and less abstract chatter.
If you look at research on extractive compression, the logic is the same: selecting important fragments is usually more reliable than rephrasing them with an abstract summary. This is especially noticeable in agentic chains, where any inaccurate generalization later breaks the plan deeper down the stack.
Impact on Business and Automation
For production, there are three direct effects. First: cheaper long runs, because you're cutting tokens without blind trimming. Second: less 'lost in the middle,' where the agent forgets a critical fact somewhere in the middle of its history. Third: easier AI integration into real processes, where the context is constantly noisy.
Teams with long workflows benefit most: development, support, auditing, document processing. Those who believe that a large context window alone replaces AI architecture lose out.
At Nahornyi AI Lab, we constantly deal with these bottlenecks: where to store invariants, what to compress extractively, and what can't be cut at all. If your agent is already burning through the budget but still losing the thread of the task, let's analyze your scenario and build an AI solution development plan so the model finally works instead of just eating tokens.