Phase 5: Agent Orchestration Patterns · 45 min · Claude Code · Python

Single-Agent Architecture Done Right

One well-harnessed agent with real state management beats a multi-agent system nobody needed to build yet.

Hiring signal: Building a single agent with real checkpointing and recovery — not defaulting to a multi-agent system for complexity's sake — shows restraint and engineering judgment that's rarer than it should be.

What you will learn

Describe the augmented LLM pattern: a single model call augmented with retrieval, tools, and memory
Identify when a single well-harnessed agent is sufficient and preferable to a multi-agent system
Design state management for a long-running single-agent task
Implement checkpointing and recovery so an interrupted task resumes instead of restarting

Introduction

Single-Agent Architecture Done Right

This lesson exists specifically because the previous lesson's five workflow patterns, and the next lesson's coverage of multi-agent orchestration, can create a false impression that more sophisticated architecture is always the more mature choice. It isn't. A large share of real production tasks are genuinely well served by a single agent, and the actual engineering skill is recognizing that clearly enough to resist adding coordination complexity a task never needed.

A team scoping a document-processing feature — read 2,000 uploaded contracts, extract key terms, flag anything unusual — starts designing a multi-agent system: a reader agent, an extraction agent, a flagging agent, an orchestrator to coordinate them. Before writing any of it, someone asks the obvious question: what would one agent, with retrieval, the right tools, and real state tracking, actually fail to do here? The honest answer is nothing — this task doesn't need specialist agents debating with each other, it needs one well-harnessed agent that can read a document, extract terms with a tool, check them against known patterns, and keep track of which of the 2,000 contracts it's already processed. They build the single-agent version. It's simpler, cheaper, easier to debug, and finishes the job just as well.

The augmented LLM: often, this is enough

The simplest production-ready agent pattern is a single model call augmented with three things: retrieval (pulling in relevant documents or facts, per c12-02's context engineering), tools (specific, scoped capabilities the agent can invoke), and memory (persisted state across turns or across a long-running task). This is not a lesser architecture waiting to be upgraded to "real" multi-agent orchestration — for a large share of real tasks, it's the correct final architecture, and the discipline this lesson teaches is recognizing when that's true instead of reaching for more moving parts by default.

The question that should come before any multi-agent design
Before designing multiple specialist agents, ask what specifically a second agent would catch or do that a well-harnessed single agent, with the right tools and a genuine validation step, wouldn't. If the honest answer is "nothing specific," the multi-agent design is adding coordination overhead — more places for context to fragment, more places for errors to propagate between agents — without buying anything the task actually needed.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers State management for long-running tasks, Checkpointing and recovery, Recovery, not just persistence, Build It — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.