Month 2 Plan

Month 2 — Agent Core: Reasoning and Tool Calling

This month you add the "intelligence" layer to the project. Careful: this month is where most people stop. The real difference shows up in Month 3, when you measure the quality of the work.

Learning core (both of you)

• LLM fundamentals (conceptual): what an LLM is, token, context window, temperature
• Neural networks and transformers (conceptual — without the math)
• Tool/function calling
• Agent loop: think → select a tool → call → observe → think again
• RAG as one of the agent's tools: embedding, chunking, similarity search

Concepts you must be able to explain

• Token and context window
• Embedding vector
• Transformer and attention (conceptual)
• Tool calling contract
• Agent loop (reasoning loop)
• System prompt
• State management
• Infinite loop risk and max steps

Project steps

1. LLM client (API key from an environment variable, never hardcoded)
2. Tool #1: RAG search (chunking + embedding + pgvector)
3. Tool #2: a deterministic simple tool (calculation / date)
4. Agent loop — each turn produces an agent_step record
5. max_steps safety limit
6. The "start task" endpoint in control-plane is opened to the end user
7. Two different system prompts are tried, and the behavioral difference is observed

Why at least two tools? With a single tool the agent doesn't decide, it just calls that tool — that's a plain function call. With two tools the question "which one, when" arises; that's the behavior that makes an agent an agent.

Division of labor (rotating)

	First half	Second half
Can	agent loop + LLM client	RAG tool + control-plane integration
Simurg	RAG tool + pgvector	agent loop + max_steps tests

Exit criteria

See roadmap.md

Content plan

• "Agent thinking" screen recording — the inside of an AI agent's head
• "What is tool calling, in 60 seconds"
• A "failure moment" clip showing the step limit kicking in after pushing it into an infinite loop
• Month 1 ↔ Month 2 comparative progress video

Retro