Brazosd is a cloud agent platform. Its agents run in the cloud and have cloud computers of their own, so they do not interfere with the computers their users are working on. Industry experts can earn revenue by publishing agents. The agents provide a conversational service without exposing the core SOP behind it. Regular users can directly use practical agents built by industry experts without first learning a collection of complicated concepts.
When a task is defined clearly enough, you can build an agent specifically for it. An agent might recognize invoices and summarize their data in a fixed format, or identify the concept behind a student's mistake and find related questions from real exams.
This series explains how to develop agents for specific business needs on Brazosd. We will use invoice recognition and exam-question retrieval to introduce the infrastructure Brazosd provides. This first article starts with one of the foundations: using Brazosd Cloud Drive to build a knowledge base.
Brazosd Cloud Drive
Cloud Drive is the medium through which agents and users exchange files. Brazosd provides several types of drives:
Organization drive: Every member of the organization can see it. Members can use it to share files within the organization.
User drive: A personal drive for each user. Only you and the agents serving you can see and manage its files.
Agent drive: The agent's own drive, belonging only to that agent. Users outside the organization cannot view its contents. When the agent serves an external user, it can read this drive but cannot write, delete, or edit files. It can modify the files only while serving a user from the organization that created it. The agent drive normally stores business-specific materials such as a knowledge base.
Channel drive: Only channel members can access it. It is generally used to keep different projects isolated.
Session drive: Every session has a persistent drive for attachments and agent deliverables. Files remain at their original paths after inactivity expiry or manual archival, until you delete them. You do not need to move them to another drive to keep them. Archival destroys the work sandbox used for computation, so save any results you want to keep to a cloud drive first.
When an agent works for a user, it uses that user's permissions for each type of drive. The agent is simply acting on the user's behalf; it does not receive a separate or higher level of access. In general, if the user can read or modify a drive, the agent can perform the same operations. If the user cannot access it, neither can the agent. The only special case is the agent drive: while serving a user outside the organization, the agent can still read its own drive, but it cannot modify its contents.
Building a knowledge base with Cloud Drive
Brazosd provides Cloud Drive, but it does not include a predefined “knowledge base” like IMA. A knowledge base is really files plus indexes. There is no single way of organizing indexes that works for every business. We should design the file and index structure around the actual requirements.
Take the task “identify a mistaken question, then find related questions from real exams.” We need an exam-paper “knowledge base” to store the papers. In practical terms, we put each paper into folders organized by year and region, then create index files that let us find questions by learning concept.
For English, we can create one index file per concept. For example, create relative clauses.txt inside the index directory. This file records which questions in which papers involve relative clauses and where those papers are stored.
When the agent needs real exam questions about relative clauses, it can read this index and locate them. The index can also record difficulty, frequency, and other details that can be used as filters.
What remains is to tell the agent how to maintain and use this knowledge base. Here is an example prompt:
# Exam Paper Knowledge Base
The knowledge base is in your agent drive. It contains two kinds of files: exam papers and indexes. For example:
Exam Papers
- 2026
- Guangdong
- Huizhou First Joint English Exam.docx
Indexes
- Relative Clauses.txt
- Non-finite Verbs.txt
Indexes are used to locate the corresponding exam papers. Each index file covers one learning concept and uses the following format:
--------------------------------------------------
Exam Papers/2026/Guangdong/Huizhou First Joint English Exam.docx Cloze question 12 Past participle as a postmodifier Difficulty 7/10
...
--------------------------------------------------
Each line records the exam paper containing a real question, its question number, the specific concept, and its difficulty for later retrieval.
# Learning Concepts
The learning concepts are: nouns, articles, pronouns, adjectives, adverbs, prepositions, verb tenses, passive voice, modal verbs, subjunctive mood, non-finite verbs, relative clauses, nominal clauses, subject clauses, object clauses, predicative clauses, appositive clauses, adverbial clauses, subject-verb agreement, inversion, emphasis, ellipsis, there-be constructions, exclamatory sentences, imperative sentences, tag questions, absolute constructions, with-complex constructions, comparatives and superlatives, word formation, and fixed collocations.
# Filing Exam Papers
When a user sends an exam paper, do the following:
1. Put the paper in the directory for its year and region. For example: Exam Papers/2026/Hubei/First Provincial Joint English Exam.docx
2. Using the learning concepts listed above, identify the concepts involved in each question and update the corresponding index files
# Retrieving Real Exam Questions
When a user sends a question, first extract the relevant learning concepts. Then find the corresponding concept indexes in the index directory of your agent drive. Filter related questions from real exams according to the user's requirements.This prompt tells the agent how the knowledge base is organized, maintained, and used. Once it becomes part of the agent's responsibility prompt, the agent can automatically file a new exam paper and update its indexes when it receives one. When it receives a question, it can use those indexes to retrieve related questions from real exams.