Why build agents without a framework?

Frameworks hide the loop, and the loop is the thing worth understanding. Once you have written it yourself, frameworks become a convenience you can evaluate rather than a dependency you cannot debug.

When should I not use an agent?

Most of the time. If the steps are known in advance, a plain workflow is cheaper, faster and more reliable. Recognising that distinction is an explicit outcome of this phase.

Phase 5 of 6 · Weeks 63–72

Phase 5 — Agentic AI

Build agents from the loop up — no framework — so you know exactly why they fail and when not to use one.

Try a lesson free — no account Am I ready for this phase?

What you'll be able to do

Build a ReAct agent from scratch, with no framework in the way
Design tool schemas that a model can actually use reliably
Handle tool errors, retries and partial failures without the agent derailing
Add planning and self-critique — and measure whether they helped
Decide when a plain workflow beats an agent, and argue the case
Connect an agent to external systems through the Model Context Protocol

What you need before you start

Assumed on day one

Phase 4 — how LLMs generate text, and their limits
Solid software engineering: APIs, JSON schemas, error handling
Comfort debugging a system whose behaviour is not deterministic

Taught here — not assumed

The agent loop — observe, reason, act — and the ReAct pattern
Tool definitions, function calling and parallel tool calls
Planning, reflection, self-critique and scratchpad patterns
Memory and context management across long-running tasks
The Model Context Protocol and multi-agent coordination

Week by week (10 weeks)

W63Agent Foundations
W64Tool Use & Function Calling
W65Planning & Reasoning
W66Memory Systems
W67Agent Frameworks I — LangGraph
W68Agent Frameworks II — Survey
W69MCP (Model Context Protocol)
W70Multi-Agent Systems
W71Agent Evaluation & Observability
W72Agent Safety

What you'll build

A single-tool ReAct agent written from scratch, no framework
An agent with five or more tools and robust error handling
A planning agent evaluated against a non-planning baseline
An agent connected to a real external system over MCP

Questions about this phase

Why build agents without a framework?: Frameworks hide the loop, and the loop is the thing worth understanding. Once you have written it yourself, frameworks become a convenience you can evaluate rather than a dependency you cannot debug.
When should I not use an agent?: Most of the time. If the steps are known in advance, a plain workflow is cheaper, faster and more reliable. Recognising that distinction is an explicit outcome of this phase.

Start with a real lesson, not a sales page.

Play through a complete lesson — predict, explore the demo, prove you've got it. No account, no card.

Try a free lesson See all 78 weeks