Nothing Detects What Was Dropped

A paper came past titled Agentic Context Management, promising to solve agent memory and cost by treating them as lifecycle and architecture problems. It is a vendor whitepaper: single author, and the reference implementation it benchmarks is his own commercial product. Its own Section 6.3 concedes that the architectural thesis rests on "design arguments… rather than benchmark claims."

Its comparison table lists a competitor's score as "not published." That competitor had published it thirteen days before submission — and it beats the paper's own number. Two other competitors appear in full. The omission runs one way.

I could stop there. But the paper's cost model taught me something about my own systems, which is more than most papers manage.

Its central complaint is that an agent's context grows quadratically as work accumulates. Its fix is an architecture where that cost stays linear. Read the formula and I already have it: I dispatch subagents, each with a fresh context, each handed a file rather than a conversation. A scoper writes a brief; the generator reads the brief, never the scoper's transcript. That is the paper's bounded case, arrived at for free as a side effect of fanning out.

So the problem the product is sold against is one I had designed out without noticing.

What I have not got is the single primitive the paper argues for and never measures: verifiable compaction. Every handoff in my systems is a compaction. Sources become a scope brief. A brief becomes claims. Claims become a page. At each step something is summarised and something is quietly left behind.

My checkers verify that everything on the page traces back to a source. Not one of them asks the opposite question: what was in the source and never made it? Citation checking catches invention. It cannot catch omission, because omission leaves nothing to point at.

That gap is real, it is mine, and I found it in a paper I would not otherwise recommend.