This idea began while I was looking at the open-source Codex and DeepSeek Harness projects. Watching these agent harnesses organize tools, tasks, and models kept bringing me back to the same question: how should they manage context?
Some models now support context windows of one million tokens, but even a million-token window is finite. The longer an agent works, the more files, conversations, and intermediate results it accumulates. It cannot keep all of them in their original form forever. Longer windows are one answer. Retrieval and summarization are others. But could language itself compress the record? Could the same material be expressed in a way that uses fewer tokens?
Chinese already has a ready-made example: Classical Chinese.
Classical Chinese is hard for modern readers. It often omits subjects and conjunctions. Sometimes even the relation between people has to be inferred from the surrounding text. The role of a character can change with its position in the sentence. These features have usually counted against it. Vernacular Chinese states more explicitly, so the reader has less work to do.
Looked at differently, Classical Chinese leaves part of that work to the reader. A person reconstructs the unstated subject or relation from context. A large language model might be able to take over this job. When context is stored, anything that can be reliably inferred could be left out. When the information is needed, the model could restore it. What used to be a burden on the human reader might become a saving in tokens.
If a passage in Classical Chinese uses only 60 or 70 percent of the tokens required by its vernacular equivalent, a fixed context window could hold considerably more material.
I wanted to see whether the numbers pointed in that direction, so I ran a very small test.
Counting the tokens
I compared four kinds of text: original Classical Chinese, the corresponding modern Chinese, ordinary English, and a deliberately compressed form of modern Chinese that remained directly readable. The last version was not code and required no separate codebook. It simply removed as much verbal excess as possible.
One sentence from Nongsang Jiyao reads:
楚汉相距于荥阳，米石至万。
A modern Chinese version is:
楚汉两军正在荥阳交战，米价上涨到每石一万钱。
A more compressed modern version is:
楚汉两军荥阳交战，米价每石涨至一万钱。
The English version used in the comparison was:
While the Chu and Han armies faced each other at Xingyang, the price of rice rose to ten thousand coins per shi.
With the Qwen3 tokenizer, the three Chinese sentences use 13, 19, and 18 tokens respectively. The English sentence uses 27.
One sentence could easily be a fluke. I then took 3,299 Classical and modern Chinese pairs from the Analects, A New Account of the Tales of the World, Nongsang Jiyao, and Zizhi Tongjian, and counted them again with four tokenizers. Classical Chinese used 25.2 to 34.2 percent fewer tokens than modern Chinese. The amount varied by tokenizer, but the direction did not. It was also the same when the four source collections were counted separately.
I next made a direct four-way comparison with 24 groups of text. Classical Chinese used 30.1 to 39.9 percent fewer tokens than ordinary modern Chinese, 6.0 to 38.4 percent fewer than English, and 6.3 to 15.0 percent fewer than the already compressed modern Chinese. Across all four tokenizers, the Classical Chinese total was the lowest.
The last range, 6.3 to 15.0 percent, is the part I find most interesting. The compressed modern version had already removed as much expendable wording as it could, yet the Classical version was shorter still. Where did the remaining saving come from? Perhaps it came from monosyllabic words and more aggressive omission. Perhaps the tokenizer was simply more familiar with certain common Classical phrases. This test cannot tell us yet.
The English results also varied widely. With some tokenizers, Classical Chinese saved a great deal. With others, it saved only a little. Compression is not determined by language alone. The tokenizer's segmentation choices change the result directly.
I am not suggesting that people start writing in Classical Chinese again
What interests me is the model's internal context. Writing meant for people should still be clear and easy to read. There is no reason to make an article resemble an ancient text just to save a few tokens. But the material a model keeps in a long context may not need to preserve its original wording.
One could imagine rewriting incoming material into a more compact natural language before it enters a long context. Names, numbers, and relations would remain. Repeated subjects and background that has already been established could be omitted. When the model needs to answer a question, it would retrieve information from this compact text. This differs from ordinary summarization because the aim is not to preserve only the gist. The aim is to preserve the original facts in shorter sentences.
The result may not look like historical Classical Chinese at all. It might sit somewhere between modern Chinese and Classical Chinese: still readable, without an external codebook, but more compressed than ordinary prose. Historical texts serve here as a reference. They show how far a natural language can be shortened while remaining language.
Twenty-four groups are far too few. The English and compressed modern versions also need independent checking to confirm that all four forms carry the same content. For now, the figures are only a starting point.
My next step would be deliberately tedious: enlarge the sample, then add the omitted subjects, function words, and relational markers back into each Classical sentence one by one. We could then see where the token count rises. That might separate savings produced by the tokenizer from those produced by the language's own way of expressing information.
It is still too early to say whether this can become a practical method for context compression. The useful thing, for now, is to keep testing the possibility.