Agentic Systems - From Zero to a Productive Assistant
This article is about agentic systems. First, we’ll cover the basics — how to start using assistants and build a proper system around them. In the second part, we’ll move on to more advanced ways of using assistants, automation pipelines, and so on.
Why You Need an Agent
The first question you need to answer honestly is: what do you need it for? The range of tasks an agent can automate is huge: from sorting documents on your PC to developing large applications. Everything is possible — the question is the cost.
Where to Start
If you’re just getting into the topic, I recommend starting with one of two tools.
The first is Codex by OpenAI. It is included in the Plus subscription. To get started, go to chat.com, select Codex in the left sidebar, download it, and sign in.
The second option is Claude by Anthropic, available at claude.ai.
How Not to Burn Money on Tokens
In the world of AI, you absolutely need to think about costs. You can’t just feed an agent gigabytes of context and hope to get a decent answer. In theory, it’s possible, but it will cost a lot.
So I suggest the following setup: the agent keeps a daily log and history, plus a reference index for quick search. The structure looks like this:
📁 Your project/
└── 📁 _agent/
├── MEMORY_INDEX.md
├── PROJECT_MEMORY.md
└── PROJECT_HISTORY.md
A separate note on folder organization: it’s better to create one large folder for each major project, and inside it create subfolders for individual tasks. Don’t mix everything into one pile.
You shouldn’t keep one chat going forever. Use it for a day, two, maybe three, then archive it and open a new one. From time to time, clean up archived chats too — they also take up memory and can become quite heavy.
Starter Prompt
Here’s the prompt I use for my personal operational assistant. Paste it into a new chat, and the agent will create the necessary structure by itself:
You are my personal operational assistant.
Work briefly, pragmatically, and without fluff.
Main task:
help with everyday work and projects,
maintain focus and priorities,
structure information,
manage project memory,
bring tasks to completion.
Token-saving mode (mandatory):
answer briefly and to the point,
do not repeat yourself,
do not give long explanations unless requested,
if there is a lot of data that we refer to often, warn that it is expensive in tokens and suggest switching to a database,
for persistent project state, prefer a database over JSON,
use JSON only as a temporary format/export.
Startup protocol (first message in a new chat):
say that you are ready to start immediately,
offer a choice:
a) first ask 3–5 questions about me and my tasks,
b) start working immediately.
Mandatory action during initialization:
create an AGENT.md file in the current working folder with these rules,
before creating a service folder, you must ask:
"Can I create _agent in this folder and store project memory there?",
if the user agrees, create:
_agent/MEMORY_INDEX.md
_agent/PROJECT_MEMORY.md
_agent/PROJECT_HISTORY.md
If question mode is selected, ask briefly:
1. Who are you and what do you do?
2. What are your 1–3 main goals for the next month?
3. What projects are currently active?
4. Where are tasks and data currently stored: files, spreadsheets, databases?
5. What is critical: speed, money, quality, or deadline?
Memory rules:
PROJECT_MEMORY.md — current state, short snapshot.
PROJECT_HISTORY.md — append-only history, only significant changes, organized by date.
MEMORY_INDEX.md — map of memory files and where to find what.
After every significant step, update PROJECT_MEMORY, add an entry to PROJECT_HISTORY, and update MEMORY_INDEX if the structure changes.
Response format:
brief summary,
next steps,
risks/blockers,
what is needed from the user, only if needed.
Focus mode:
if there are too many tasks, split them into Critical / Important / Optional,
limit active WIP to 3–5 items,
track repeated actions as habits, not chaotic duplicates.
Limitations:
do not invent facts, statuses, sources, or progress,
separate facts from hypotheses,
do not stretch out responses or use motivational fluff.
Which Model to Use
This is simple: flagship models are expensive, and most of the time you don’t need them.
For OpenAI, GPT-4o, GPT o3 mini, or Codex are good choices for everyday tasks — they cover most use cases. GPT-4.5 and GPT-5 should be reserved for rare cases when the task is genuinely serious. In that case, it’s better to go to the web version and conduct deep research there.
With Anthropic, the logic is the same. Claude Sonnet is great for routine work — fast and inexpensive. Use Opus only when you need real depth: complex analysis, large documents, non-trivial tasks.
If you need a detailed answer on a complex topic, create the prompt together with the agent, embed your preferences and current context into it, and then take that prompt to the web version for deep research. The answer will be much more accurate and personalized.
Connecting External Tools
An agent doesn’t have to work in isolation. You can connect Notion or Obsidian as a knowledge base, GitHub for working with code and repositories, and various task-tracking systems. I chose to develop my own trackers for myself — no unnecessary dependencies, and exactly the functionality I need.
The Main Point
An agent is not a magic button — it is a tool. The better you build your memory system, folder structure, and habit of archiving chats, the cheaper and more accurate it will be. Start with one project, customize the prompt for yourself, choose the model based on the task, and don’t pay for flagship models when a lightweight model can handle the job.
What’s Coming in Part Two
In the next article, we’ll look at more complex architectures. We’ll talk about multi-agent pipelines, agentic applications like OpenClaw, Hermes, and others. I’ll explain how to set up a system where different tasks are handled by different models — so you pay for tokens, not subscriptions. This is often more cost-effective, depending on the tasks.
We’ll also cover automation through n8n, talk about Cockpit, and discuss current trends in AI.
Top comments (0)