// Topic

AI Agent Architecture

The patterns, trade-offs, and failure modes behind production AI agents. From the agent loop and tool calls to multi-agent design and error handling.

An agent is a loop. It perceives, decides, acts, and remembers, then does it again until the work is done or it gives up. Everything hard about agents lives in how you wire that loop and what you do when a step fails.

I have spent two years watching agents break in production and mapping the design space across real deployments. The posts here cover the anatomy of the loop, the choice between one agent and many, the harness that holds it together, and the errors nobody warns you about.

Foundations

What an agent is, how the loop runs, and the shape of the design space.

A Taxonomy of AI Agents That Actually Explains What You Are Building

Most AI agent taxonomies are either too academic or too vague to be useful. Here is the classification I use when I need to decide what kind of agent to build.

8 min read ai agents architecture

When to Build an Agent and When to Build a Smarter Assistant

The difference between an AI agent and a smart assistant comes down to one thing: who drives the loop.

6 min read ai agents ai architecture agent design

The Anatomy of an Agent Loop: Perceive, Think, Act, Remember

The agent loop is not one thing. It is four distinct phases that run in sequence, and understanding each one is how you debug what breaks.

8 min read ai agents agent architecture loop design

The Agent Design Space: A Map of What Engineers Are Actually Building

After surveying production agents across industries, the design space clusters into patterns. Here is what I found.

8 min read ai agents architecture

Single vs multi-agent

When more agents help and what the coordination costs you.

Multi-Agent vs Single-Agent Systems: The Real Trade-offs

The decision between one agent and many is not about capability. It is about failure modes, latency, and operational complexity.

6 min read ai agents multi-agent systems agent architecture

Shared Memory vs Isolated Memory in Multi-Agent Workflows

How to choose between shared and isolated memory architectures for multi-agent systems, with trade-offs from production deployments.

8 min read ai-agents multi-agent agent-memory

Tools and infrastructure

The plumbing that lets an agent act reliably on the world.

The Model Context Protocol (MCP) Explained: A Universal Language for AI Tools

Model Context Protocol (MCP) is the new standard for connecting AI models to data sources and tools. Here is why it matters, how it works, and why it is the missing link for agentic infrastructure.

8 min read ai agents infrastructure

Agent Harnesses: the Infrastructure Layer Your Llm Agent Actually Needs

Every production AI agent needs a harness. Here is what one contains, why frameworks often are not enough, and how to build the layer that actually determines reliability.

10 min read ai agents infrastructure

Structured Outputs with LLMs: JSON Mode, Function Calling, and When to Use Each

JSON mode, function calling, and structured outputs solve different problems. Here's when each one actually makes sense and what they cost you.

12 min read ai llm infrastructure

Reliability in production

The failure modes that break real deployments and how to handle them.

Why AI Agents Keep Failing in Production and What the Field Is Doing About It

I have spent two years watching agents fail in production. Here is what I keep seeing and what the field is starting to do about it.

6 min read ai agents production reliability

What Nobody Tells You About Error Handling in Production AI Agents

Hard-won lessons from running AI agents in production: the error patterns that actually break systems, and the patterns that fix them.

8 min read ai devtools backend

Why Your Agent Remembers the Wrong Thing: Memory Attribution Failures

Memory attribution failures cause AI agents to act on outdated or misassigned context. Here is what I found examining the failure patterns.

8 min read ai-agents agent-memory production-errors

Benchmarks

Head-to-head data on agents doing real work.

Beyond the Chatbox: Benchmarking Claude Code vs. Gemini CLI for Autonomous Repository Refactoring

I ran a head-to-head benchmark of Claude Code and Gemini CLI on an M2 Air (16GB), testing their ability to autonomously refactor legacy code and resolve race conditions.

7 min read agentic-cli benchmarking developer-productivity

Frequently asked questions

What is an agent loop?

An agent loop is the repeating cycle an agent runs: perceive the current state, decide the next step, act with a tool, and record the result. The loop continues until the task is complete or a stop condition fires. Almost all agent behavior is explained by how this loop is structured.

When should I use multiple agents instead of one?

Reach for multiple agents when the work splits into clearly separate roles that each need their own context and tools. The cost is coordination, latency, and new failure modes. For most tasks a single well-instrumented agent is simpler and more reliable.

Why do AI agents fail in production?

The common causes are brittle tool schemas, missing error handling, context that grows until it poisons reasoning, and no observability into what the agent did. The production-errors and why-agents-fail posts cover each pattern and the fixes.

What is the Model Context Protocol (MCP)?

MCP is a standard for connecting AI models to external tools and data sources through a common interface. It lets you wire the same tool into many agents without rewriting integrations each time.

Need a technical writer who gets this?

I write for DevTools and B2B SaaS companies. Let's talk.