Agentology · Knowledge of agents
of an answer,
to an action
Know your agent from the inside; The target, the tool, the memory and the loop that checks the result. This is the starting point of your learning path in Nima creations.
Discover the path
Why try this product?
See your agent from the inside, not from the definition.
Agentology is a 10-stop path that, instead of memorizing jargon, exposes how agents work: how they make decisions, call tools, recover from errors, and collaborate.
- 6 interactive simulatorsRun your own Agent Ring, RAG, Multi-Agent System, Token Calculator and Security Lab.
- Glossary + 20 question testRead each concept with an example and measure yourself at the end.
- Suggested path based on your problemFirst, determine what your problems are; Then enter the repair workshop.
Do you want to learn more and differently?
First, register your profile and email address here to get full access to this educational content, simulators, and games — once for all builds.
No spam; Only the news of new creations and courses. Do you have an account? enter
First, know the problem.
Then make an agent.
Not all jobs need an agent. A clear request may be resolved with an answer; A fixed process with workflow; And an open problem with a cycle of decision and tools.
Agentology is part of Nima's creations: short lessons, touchable simulators, and a path from prompts to controlling system behavior.
For your problem, which path?
Choose a scenario.A clear prompt is enough.
When you just want a text of the available information, start with the specified command and evaluate the output.
Prompt engineering workshop →Continue your path with practice.
Igentology, step by step
13 main lessons and Igentology simulators. Open each lesson and go through the examples. The data and costs of the simulators are educational and not the daily price of the services.
Why is the agent you are building not working?
There is a repeating pattern: you watch a tutorial, build the agent, the first test works great, but on the second or third task everything falls apart! The agent is either going around in an endless loop, or hallucinating, or generating a hefty bill. The reason is not that the code is wrong; Not knowing is that What exactly is going on under the hood of the agent?
Runaway Loop
The stopping condition of the agent is weak. Instead of solving the problem, the model spins endlessly and consumes tokens.
Hallucination
You have assumed that the model knows your private information, while the unprovided private information is not necessarily in the input or available tools of the model!
Token Inflation
At each round of the loop, all previous history is fed back to the model, and the cumulative input consumption grows faster than the number of rounds.
From neural network to linguistic model (LLM)
Traditional programming means writing precise rules: "if this, then do that". Artificial intelligence reversed this relationship; Instead of a rule, you show thousands of examples so that the machine can make its own pattern. In 2017 architecture Transformer with mechanism Attention It was introduced that it processes the whole words of the sentence at once.
📌 The basic difference between traditional programming and AI:
In traditional code, any unexpected input crashes the system. In the linguistic model, the output is a "high-quality statistical prediction for the next token" based on the word's attention weight.
⚡ What is the Attention mechanism?
The old models went forward word for word. Transformer looks at the whole sentence at once and calculates how dependent each word is on other words.
The vital trinity: chatbot, workflow or agent?
The biggest misunderstanding in the world of artificial intelligence is here! Most of the tasks that people create agents for can be solved much cheaper, faster and 100% more predictable with a simple workflow.
Chatbot
He waits for the question, answers, that's it. Can't do anything outside the chat box.
- Suitable for answering common and simple questions
- Very low cost and instant response speed
- It is not possible to do practical work and change the file or database
- There is no multi-stage decision-making and trial and error
Workflow
You have already determined the path step by step (first A, then B, then C). The model is not a route decision maker.
- 100% predictable and without logical error
- Controlled and fully optimized token cost
- Very fast and stable in enterprise environments
- It is not flexible for uncertain and unknown scenarios
Agent (AI Agent)
"Target" takes, not the question. He decides which tool to take and what is the next step.
- Solving multi-stage, uncertain and complex problems
- Self-correction, debugging and repetition until reaching the result
- Variable cost and requires strict guardrails
- Requires human supervision (Human-in-the-Loop)
🧪 Diagnostic tool: Which one does your project really need?
Choosing between chatbot, workflow and agent is the biggest decision in cost optimization and sustainability of your product.
Agent Loop - The beating heart
The heart of every agent is a 4-step loop: thinks, act, observes the result and Decide/Check Stop. This cycle is repeated until the job is done.
Tools and Function Calling
The agent does not get anywhere by thinking alone; Must be able to act in the real world (search the web, read a database, run code, send email). But how does a model that produces only text call the tool? Answer: Function Calling in JSON format.
📦 Standard model output format (JSON):
does not "execute" the instrument model; The model only writes a JSON structure. Your external program reads it, executes it, and returns the result:
⚠️ The golden rule of the number of tools:
As the number of tools available to the model increases, the probability of model confusion and wrong tool selection increases. Always start small (1-3 essential tools).
Memory, context window and token economy
Linguistic models by nature Stateless are Every time you talk to them, it's like they just woke up! Therefore, the entire history must be resubmitted. In Agents, each round of the loop enlarges the history, and the cost of tokens increases as the history accumulates.
🎛️ Change the agent execution parameters:
Embedding, vector database and RAG
How to find only the relevant part among thousands of pages of information? Sentences are converted into numbers (vector / Embedding). Sentences with close meaning (such as "the cat slept on the carpet" and "Peesi da lee on the carpet"), their distance in the vector space is close to zero!
🔍 Semantic search and retrieval:
...
MCP — Standard Tools Port (Model Context Protocol)
Previously, to connect each tool (Gmail, Drive, Postgres) you had to write specific code for the same model. Protocol MCP It is a universal open standard so that any tool server written once can be used by all agents and models without code changes.
Asymmetric cable connections
A recode for OpenAI, another for Claude and another for LangChain for each tool!
Universal USB port
Write the MCP server for your GitHub or database once; All clients (Cursor, Claude Desktop, Antigravity) connect instantly.
MCP Client & Server
The client sends the prompt, the MCP server returns the list of methods and resources (Resources & Tools) by default.
The four main architectures of genetics
ReAct
Reasoning + Acting. Suitable for short and quick tasks.
Plan & Execute
First, he writes the complete checklist of the route and then executes it one by one so as not to deviate from the goal.
Reflection
Production ⟵ review ⟵ automatic rewriting to bring the quality of text and code to the highest level.
Dynamic Routing
Intelligent routing of the user's request to the relevant agent or specialist workflow.
Multi-Agent Systems
When you have an agent with 20 tools and a 10-page manual, you're asking one person to be versatile, and the result is poor. Solution: role specialization with several centralized agents.
👥 Selection of team scenario:
Frameworks and SDKs
LangGraph and CrewAI
LangGraph: An option for when the workflow has branching, complex conditions, and stop/resume capabilities.
CrewAI: Focused on multi-agent systems with role segregation and automated team collaboration.
Anthropic, Google GenAI & OpenAI
provide interfaces to use the capabilities of the constructor model; But be careful Vendor Lock-in Be so that you can easily move the models in the future.
Security, Prompt Injection and Guardrails
When the chatbot makes a mistake, you see wrong text and get rejected; But when the agent makes a mistake, a file is deleted, code is executed, or data is leaked!
...
...
...
Real applications in industry
Automated coding and testing
An agent that writes code, executes unit tests, debugs in case of errors, and commits when the tests pass.
Email triage and meetings
Categorize incoming messages, extract meeting tasks, and draft responses in the human approval queue.
Analysis of financial documents and RAG
Reading hundreds of pages of annual financial reports, crunching the numbers and extracting opportunities and risks by checking the accuracy of the output.
Additional study: Agent and Workflow Patterns — Anthropic
From learning to implementation