🏗️ From QA Bot to Task Agent: An Architecture Guide

TL;DR: Stop building chatbots that just answer questions. Start building task agents that actually get work done.

This guide explains the architectural shift from monolithic QA bots to Task Agents using Static Rules, Dynamic Skills, and Deterministic Hooks—with concrete code examples and open-source references.

Illustration of a task-agent system architecture showing interaction between different components

1. The Core Shift: QA Bot → Task Agent

Most AI systems today are still context-stuffed QA bots:

• They answer questions well

• They hallucinate under pressure

• They lack guarantees around execution, safety, and consistency

💡 The key insight: Don't scale context. Structure it.

2. The Three-Layer Architecture

🧱 1. Static Context — Rules (Always On)

•Mental model: Employee handbook
•Always loaded
•Defines identity, coding standards, behavioral constraints
•Prevents hallucinations and style drift
•Small, stable, human-editable

🛠️ 2. Dynamic Context — Skills (On Demand)

•Mental model: Toolbox
•Loaded only when needed
•Each skill is a self-contained capability
•Keeps the context window clean

⚓ 3. Deterministic Hooks — Guardrails

•Mental model: Security + Compliance layer
•Not probabilistic
•Runs before / after LLM reasoning
•Enforces rules that must never fail

3. Recommended Project Structure

my-task-agent/
├── .cursorrules
├── main.py
├── tools/
│   └── linear_mcp.py
└── README.md

4. Static Context Example: .cursorrules

# ROLE
You are a Senior Python Engineer focused on production-grade systems.

# RULES
- NEVER use print() for debugging
- ALWAYS type-hint functions
- Propose a plan if touching >3 files

# BEHAVIOR
- Be concise
- Ask clarifying questions if needed

Reference: https://github.com/PatrickJS/awesome-cursorrules

5. Dynamic Skill Example (MCP)

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("DevTools")

@mcp.tool()
def create_linear_ticket(title: str, priority: str = "low") -> str:
    ticket_id = f"LIN-{hash(title) % 10000}"
    return f"Created ticket {ticket_id} with priority={priority}"

if __name__ == "__main__":
    mcp.run()

Reference: https://github.com/modelcontextprotocol/python-sdk

6. Deterministic Hook Example

def compliance_check_hook(state):
    user_input = state["messages"][-1].content.lower()
    if "password" in user_input or "api_key" in user_input:
        return {"error": "Security violation detected"}
    return agent_node(state)

Reference: https://langchain-ai.github.io/langgraph/

If your agent only answers questions, it's a chatbot. If it reliably executes work, it's a Task Agent.

AI Architecture