Two years ago, I didn’t set out to learn AI. I just started using AI coding assistants.

At the time, it felt like a better autocomplete. But it slowly turned into something much bigger.

Over the last two years, my work with AI went through this evolution:

AI assistants → repo-level AI tools → LLM APIs → embeddings → RAG → tool calling → skills → agents → workflow engines → memory systems → multi-agent architectures → AI platforms

This post is my story of that journey, and how the AI stack actually looks in practice.

---

AI Assistants

The first change was speed.

search → read → try → fail became ask → generate → run → fix

Learning became faster.

---

Repo-level AI

Tools started to understand the whole project, not just one file.

AI began helping with architecture.

---

LLM APIs

Calling models directly taught me:

model + tokens + context = AI

Everything else is engineering.

---

Embeddings

Embeddings let AI use my data.

Text → vector → search → context

This enabled real applications.

---

RAG

RAG made AI useful for knowledge bases, docs, and internal tools.

But it still only answered questions.

---

Tool calling

Now the model could do work.

Query DB Call API Read file Run job

But complexity grew fast.

---

Skills

Skills reduced tokens and made behavior reusable.

Instead of repeating prompts, I built reusable capabilities.

---

Agents

Agents combine skills + tools + memory.

Now AI could act.

But agents need orchestration.

---

Workflow engines

Real systems need:

queues retries state scheduling

This is where AI becomes backend engineering.

---

Memory

Agents need long‑term state.

history vector memory events facts

Memory makes behavior stable.

---

Multi‑agent

Real systems need many agents.

planner worker tool memory

Now AI looks like distributed systems.

---

Platforms

The next step is platforms.

Not one agent.

Many agents, shared tools, shared memory.

---

The stack

Product Platform Multi-agent Workflow Agents Skills Tools RAG Embeddings LLM API Model

---

What I learned

The future is not prompts.

The future is systems.

Cost, control, and reliability force better architecture.

Agents require backend, not less backend.

Platforms will win.

---

Conclusion

Two years ago I started using AI assistants.

Today I see a whole new software stack.

And we are still at the beginning.