Progressive disclosure for MCP/API
Progressive Disclosure для MCP/API
B · Useful, limited Short first-hand post by Rinat Abdullin, a widely followed practitioner, reporting his own experiment and a Cloudflare comparison, with few measured figures.
Summary based on the full text·Original: about 2 minutes to read
In short
A Telegram post on the channel "LLM под капотом" by Rinat Abdullin, published 9 October 2026, with 4.78K views. He describes the twelfth iteration of his "Personal OS" experiment, where the whole interface for agents and people working with versioned documents shrank to two commands. The agent starts a session, reads the current SDK and writes Python code that runs in a server-side sandbox. He links this to Cloudflare's Code Mode and argues that company processes can be encoded in such an SDK without growing the model's context.
Key numbers
- 2commands in the whole agent interface: start a session, run a Python script
- 2xreduction in agents' idle wandering time after about five SDK optimisation experiments
- 2,549Cloudflare tools, described by 1M-2M tokens, packed behind two tools
- ~1,000xtoken reduction for typical agent tasks reported for Cloudflare's Code Mode approach
As reported by the source.
Russian·Web page·opens t.me
What it reports
- Interface:
start "task name"opens a session and returns the current SDK description;run "Python script"executes code on the server in a sandbox with the user's permissions. - Document and version handling was packed into the SDK, so tool improvements no longer require redeploying API servers and all clients.
- About five SDK optimisation experiments with more convenient functions halved the time agents spent wandering.
- Cloudflare had to cover 1M-2M tokens describing 2,549 AI tools; with progressive disclosure, Code Mode and an SDK it cut token use for typical agent tasks by about 1,000 times and reduced the tools to two.
- His claim: the context stays at about 1,000 tokens even if the SDK grows to 2,500 tools.
- Processes such as creating invoices, where to store them, where to pull statistics and how to version them can ship as SDK code, picked up by new agent sessions on the fly.
- The headline claims token savings of about 99%.
How they got it
A personal engineering experiment, reported informally in a short channel post. The Cloudflare figures are second-hand, from a link a contact shared.
Limits
- No task set, baseline or measurement method is given for the 2x figure.
- The Cloudflare numbers are quoted, not verified by the author.
- The headline (~99%) and the body (~1,000x for Cloudflare) describe different measures.
Why it matters for companies
Tool catalogues for agents tend to grow until their descriptions crowd out the work. Exposing one SDK behind a small entry point, with details loaded only when needed, keeps cost and context flat as capabilities grow. It also changes how process knowledge is maintained: rules for invoices or reporting live as versioned code that every new agent session picks up. The trade-off is that running generated code needs a real sandbox and a permission model.
Key passages in translation
Russian original
упаковал весь интерфейс для работы агентов и людей с версионированными документами (аналог своего GitHub + markdown) в две команды для CLI/MCP
I packed the entire interface for agents and people working with versioned documents (an equivalent of my own GitHub + markdown) into two commands for CLI/MCP.
Russian original
Это позволило перестать, наконец, делать новый деплой серверов API-шек и всех клиентов на каждое улучшение тулзов.
This finally let me stop redeploying the API servers and all the clients for every tool improvement.
Russian original
Я смог прогнать пяток экспериментов по оптимизации SDK (более удобные функции) и сократил время праздношатания агентов в два раза.
I was able to run about five experiments optimising the SDK (more convenient functions) and cut the agents' idle wandering time in half.
Russian original
Теперь же можно в этот SDK вшивать процессы по работе с информацией в компании не словами, а кодом
Now the company's information-handling processes can be built into this SDK as code rather than as words.
Russian original
Благодаря progressive disclosure, он как был в 1000 токенов, так и останется, даже если SDK вырастет до двух с половиной тысяч инструментов
Thanks to progressive disclosure, it stays at 1,000 tokens, as it was, even if the SDK grows to two and a half thousand tools.
Summary based on the full text. Read the original for the full argument and data.Read the original on t.me
- Tags
- Cost of AI