A chain of agents condensed 32 GB of company data into a single report, opening 30,000 tickets along the way
How do you get an AI to work through 32 GB of company documents when that much data won't fit in its context? The author of the "LLM under the hood" channel split the data into chunks and handed them to agents using a Map Reduce scheme. Tickets, as in an ordinary team, kept the agents' work coordinated. The system opened more than 30,000 tickets, ran for more than a day and produced a short report for executives. The practical takeaway: keep agent memory in readable files, so you can check it.
AI-processed from TG: @llm_under_hood (LLM под капотом); edited by Hamidun News
Imagine being handed 32 GB of company exports (text mixed with JSON) and asked to tell management, in a single pass, what to fix in the business first. No model's context window can hold that much, and an ordinary chat is useless here. The author of the Telegram channel "LLM under the hood" solved this task last week. He built a chain of agents that opened more than 30,000 tickets, worked for more than a day and condensed everything into a compact report, as if very expensive consultants had written it. If you build agent systems for real work rather than for demos, the scheme is worth a look.
How it worked in practice
The agents ran sequentially, and each had its own step. The first steps were handled by Fable 5.2 in research mode. Its task was to find best practices for a company of this type and compile them into a report with links to cases and studies. This was the "ideal reference" against which reality was later compared.
Agents running on Codex came next. They combed through the entire data set and collected cases, built a map of the company's processes and performed a gap analysis: how the real process differs from the ideal one. The output was a prioritized report for executives: what to fix first and why.
Before starting, each agent loaded a project description and shared context from the "Company Brain" (the author's name for his knowledge base) at the very beginning of its context window. He compares this to loading a kernel: the agent first sees the big picture, then goes off to do its own task.
How to fit 32 GB into windows a few pages long
The author built an optimized format for representing the data and split the work according to the Map Reduce principle. Put simply, each agent takes its own chunk of the data set, and the results are then merged. For clarity, here is an example: one agent reads sales documents and leaves a note saying "contracts diverge from how the process is described in the regulations"; another picks up that note and looks for similar discrepancies in procurement. This illustrates the principle and is not an example from the case itself.
The data lived in Company Brain as files, modeled on a personal operating system. The agent does not read everything in sequence. It moves from the general to the specific (progressive disclosure) and uses the scripts docs/tools/tree.py and outline.py, which show the structure of folders and documents. To coordinate work and pass context between agents, the author added a new tool for working with tickets.
Vector databases and similar black boxes found no place in the scheme. The pipeline and its reasoning had to be tuned constantly to the priorities of the executives who would read the report. To make that possible, the author restructured the analysis process itself so it could be optimized. Tickets and text files can be read, edited and moved around by hand. You can't debug the contents of a vector index that way.
What came out of it and what's next
The final output turned out to be interesting. Besides a prioritized list with justification and links to sources, the agent advised sending a short list of questions to specific people, collecting their answers and loading them back in. In other words, the machine itself realized it lacked data and worked out exactly whom to ask for it.
The author sees the next step as a change of perspective. Instead of one "expensive consultant," he wants to run the data set through several: the most insightful HR specialist, the most paranoid security expert, a top expert in operations. For each, a methodology can be assembled from the best books and studies.
According to the author, you can assemble the methodologies yourself. It is enough to point ChatGPT or Claude at public articles and videos from strong companies, from OpenAI and Anthropic to integrators, and ask it to organize them into a knowledge base. All of this is openly available.
Caveats. This is a personal account, not a study. The author does not give the cost of the run, does not compare the report with the work of human consultants and does not explain how he checked the accuracy of the conclusions. He himself writes "let's see what happens next": executives' reaction to the report is still unknown.
The practical takeaway for developers and product managers is this. If your data exceeds the context window, split the work with Map Reduce, keep memory in readable files, coordinate agents through tickets and give each one shared context at the start. And think ahead about how a human will be able to check the intermediate steps.
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Infostart connects Direct, Metrica and Wordstat to ChatGPT: 6 commands, shared access
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
Key Questions Answered
How it worked in practice?
The agents ran sequentially, and each had its own step. The first steps were handled by Fable 5.2 in research mode. Its task was to find best practices for a company of this type and compile them into a report with links to cases and studies. This was the "ideal reference" against which reality was later compared. Agents running on Codex came next.
How to fit 32 GB into windows a few pages long?
The author built an optimized format for representing the data and split the work according to the Map Reduce principle. Put simply, each agent takes its own chunk of the data set, and the results are then merged. For clarity, here is an example: one agent reads sales documents and leaves a note saying "contracts diverge from how the process is described in the regulations"; another picks up that note and looks for similar discrepancies in procurement.
What came out of it and what's next?
The final output turned out to be interesting. Besides a prioritized list with justification and links to sources, the agent advised sending a short list of questions to specific people, collecting their answers and loading them back in. In other words, the machine itself realized it lacked data and worked out exactly whom to ask for it. The author sees the next step as a change of perspective.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.