Let's start with the task
We determine the useful result, the initial data and the boundaries of the work. The choice of model follows them.
Agent Systems Laboratory
How does an idea become a working system? We explore the architecture of AI agents, test solutions and share best practices - along with errors and limitations.
We understand the structure of the system: what problem it solves, where it gets the data from, what actions it performs and how the result is checked.
We determine the useful result, the initial data and the boundaries of the work. The choice of model follows them.
Linking roles, memory and tools. We understand why it is designed this way and what can be replaced.
We compare the answer with the source. We analyze failures, stopping conditions and decisions that remain up to the person.
We show the progress of the research and the basis for the conclusions. We discuss where the approach is useful and where it adds complexity.
Code is a means of implementation. Understanding the process helps you choose products wisely.
Towards building your agentWe are gathering an assistant for the daily tasks of the sales department:revenue and dynamics analysis, analysis of requests based on client correspondenceand preparation of reports. The agent finds the client's needs, application status and next step, supporting the answer with calculations and quotes.
We disclose the sales forecast viabasic estimate of expected demandand replenishment of the warehouse. The agent finds a ready-made calculation based on sales history in analytical documents and helps check its assumptions.
Ten steps lead from the first calculation to event reporting.Additional voice interfaceallows you to dictate a request, confirm the recognized text and listen to the response.
We connect the sales table. The agent selects the desired calculation and displays the result along with the source.
We set the request and response fields. Dates, names and indicators are brought into a single form and are checked.
The agent sends a request to the catalogue, sales or dynamics calculation. Each route ends with a clear result.
The system saves the manager, period and indicator. A short clarification continues the analysis taking into account the previous request.
Each action receives an identifier. Test cases help you compare results after changes.
We analyze applications based on client correspondence and documents. The agent finds the terms and agreements, shows the exact quote and source.
We add sales, balances and documents from different sources. Access rights determine the actions an agent can take.
We collect calculations and sources into one sequence. The result opens as a ready-made report in the browser.
We configure roles, confirmation of significant actions and backup. The working instance is validated by a user request.
A new file or schedule starts the report. The queue stores the event, retries the operation after a failure, and prevents duplicates.
Dictate a question about sales, requests or demand forecast, check the recognized text and confirm sending. In the agent project, the voice interface transmits the request to the same tools and allows you to listen to the ready response after setting up recognition and voiceover.
Analysis of voice interactionExample requests
01 / Sales
“What is the total revenue for January 2026 based on sales data?”
02 / Application Analytics · RAG
“What documents are requested in application req-0003 and how did the application end according to the correspondence?”
03 / Sales forecast
“How do balances compare with the forecasted demand for the delivery date based on August data and what assumptions are included in the calculation?”
MeaningRecognize speech and show query text
ControlConfirm the text before sending it to the agent
ResultShow the result with the source and voice the answer
From designing the system to checking the result. Open a single question or expand all answers at once.
Agents Effect Lab is a laboratory of agent systems. Here we explore the architecture of AI agents, analyze the tools and share design experience. The basis for decisions is important in the materials: what task is set, how the result is confirmed, and where the limitations remain. The central system in focus is the agent - assistant of the sales department.
An agent-assistant to the sales department helps analyze revenue and dynamics, sort out requests from client correspondence and prepare reports. To forecast sales, he uses the calculation of expected demand and replenishment needs available in the documents. The voice interface allows you to dictate a request, confirm the text and listen to the answer. Calculations and conclusions can be verified with sources.
Chat is a way of interaction, and an agent is a way of organizing the operation of the system. The agent can choose the next step and access tools within the specified rules. At the same time, the chat can also have access to search and files. The difference is determined by the independence of actions and the structure of the process, and not by the type of correspondence window.
The architecture connects the task, data, roles, tools and verification of the result. It allows you to determine in advance what the system is allowed to do and when it should stop. Without these solutions, replacing a model or tool will rarely eliminate the cause of the error. Therefore, in the laboratory, we first consider the design of the process, then the means of its implementation.
The agent calculates sales by manager and period, compares indicators and finds information on the application: customer need, conditions, status, reasons for refusal and the next step. The analysis is based on available correspondence and ready-made analytical sections. To plan sales, the agent helps to study the assessment of demand and warehouse replenishment with calculation assumptions. The manager checks the reasons and makes a decision.
The sales forecast is represented by a base estimate of expected demand and replenishment needs. The completed calculation uses August average daily sales, delivery time, and seven days inventory; the agent finds it in analytical documents and helps to sort out the assumptions. This is a calculation based on a demo set with a short history. Applying it to your business requires up-to-date data and testing assumptions.
To understand the architecture, it is more important to be able to describe the task, data flow and result criteria. Code becomes essential when implementing custom logic, connecting systems, and testing behavior. Visual tools can make assembly easier, but access, status, and error issues remain regardless of the implementation method chosen.
The choice starts with requirements: what sources are available, what actions are needed, where the data is stored, and how errors are checked. Tool limitations, operating costs, and replacement feasibility are then assessed. The popularity of a product does not in itself confirm suitability. In the blog we look at n8n, LangGraph, MCP and other tools through their role in the system.
The project includes a sales table, catalog, warehouse balances, documents and client correspondence. The extended demo set contains 50 requests: their correspondence and ready-made analytical sections are available through document search. For arbitrary calculations based on new structured data and changes to orders in CRM, a separate connection will be required. Access to correspondence takes into account the role and responsible manager.
The vector database stores numerical representations of objects and allows you to find close distances based on the selected measure. When working with text, this helps to look for semantic matches, even when the wording is different. Proximity does not prove the truth of a statement. Filters and full text search may be required for exact codes, dates and names.
RAG is an approach in which the system first finds information in the sources, then passes it on to the model to prepare a response. This helps link the response to available documents. The result depends on the quality of the search, selection of fragments and checking of links. The mere presence of a RAG does not guarantee accuracy and does not exclude unconfirmed conclusions.
Context - information available to the model in a specific request. Memory is stored data that the system can select for subsequent access. These may be agreed upon conditions of the task or the results of previous actions. Memory requires rules of selection, retention and deletion: the accumulation of all correspondence often adds outdated information and contradictions.
Validation compares statements with problem sources and criteria. Every significant conclusion must have an accessible basis; calculations require separate verification. It is useful to distinguish between facts, assumptions, and data gaps. The assertive tone of the response and the presence of references do not, in and of themselves, prove that the source actually supports the conclusion.
Correct behavior is to indicate a lack of data, clarify the request, or stop. A repeated search is useful if the source or selection conditions change. Endless repetition does not create evidence. The architecture predetermines the limit of attempts and the method of transferring the task to a person so that the lack of an answer does not turn into a fictitious conclusion.
Restrictions are set in connections and services: minimum rights, permitted operations, area of available data. Significant actions may require human confirmation. A textual request from a model to comply with the rules does not replace technical control. The content of external documents is treated as data because it may contain attempts to override instructions.
Multiple roles are useful if their responsibilities and review grounds are truly different. For example, information retrieval and assertion checking can be separated if there are common outcome criteria. Each additional role increases the cost and complexity of approval. The agreement of several models is not independent evidence if all rely on the same error.
In the local app, speech is converted to text. The user verifies the recognized command and confirms sending, after which the agent executes the request with the same access rights as in the text chat. The ready answer can be voiced. Speech recognition and synthesis require configuration; the voice interface is described as an agent project capability.
Cost and latency are affected by the model, context size, number of hits, search, and external tools. Retries and additional roles also waste resources. It is useful to compare solutions based on the cost of a proven result for a specific task. The low cost of an individual response says little about the overall costs if the response has to be corrected multiple times.
Ten steps sequentially assemble an agent - an assistant to the sales department: calculations, data structure, selection of tools, memory, log, document search and external sources. Then the report, access rights, event triggering and notifications are added. The expanded set provides materials for analyzing applications and assessing demand. Voice control is connected with an additional step: recognition, confirmation of the text and voicing the answer.
Start with one limited task: determine the expected result, available sources and method of verification. Record actions that require your confirmation. Then match the problem with the architectural solutions from the blog and the lab materials map. This order helps to discuss the specific system and its limitations before choosing tools.
Let's continue the research
Watch the ten steps of assembling an agent - a sales assistant. Each step adds a new working opportunity.