Last Tuesday I told Claude: "Find my last three email threads with the automotive client, check Google Drive for the latest SAP migration status deck, cross-reference the open items in Notion, and draft a progress update email to the project sponsor. Use formal German."

Then I went to make coffee.

When I came back, the email was sitting in the compose window. It had pulled the right threads, found the right deck, identified the three open items from Notion that actually mattered, and written a status update in the exact tone I use with that client. I read it, changed one word, and sent it.

That's not a chatbot answering a question. That's an agent doing work.

The word "agent" is everywhere right now, and most of the conversation is hype. But underneath the noise, something real has changed in how Claude operates — and if you've been following this series, you've already seen the building blocks. Memory gives Claude context. Projects give it boundaries. Artifacts let it build things. Extended Thinking lets it reason. Agents are what happens when you put all of that together and point it at a real task.

Most of the architectural thinking behind this comes from Anthropic's own research — particularly their "Building Effective Agents" piece and their learning platform. What follows is how those ideas actually play out when you're a consultant using Claude every day.

Workflows vs. Agents: The Distinction That Matters

There's an important difference that most people miss when they hear "AI agent." Anthropic draws a clean line between two things:

Workflows are sequences where the steps are predefined. You design the path, Claude follows it. First do this, then check that, then produce this output. Predictable, consistent, and you control the logic.

Agents are systems where Claude decides the steps itself. You give it a goal, and it figures out which tools to use, in what order, and when to stop. Flexible, adaptive, and you trust the model to navigate.

The key insight from Anthropic's research: the most successful implementations use the simplest approach that works. Don't reach for an agent when a workflow will do. Don't build a workflow when a single prompt is enough.

The Patterns You're Already Using

Here's what's interesting. If you've followed this series and configured Claude with memory, projects, and connectors, you're already using agentic patterns — you just might not have the vocabulary for them yet.

That email I described at the top? That was an orchestrator-workers pattern. Claude broke the task into sub-tasks (search email, search Drive, check Notion), executed each one, reasoned about the combined results, and produced the deliverable. I didn't design that workflow. I gave it a goal and it figured out the steps.

When to Let Claude Drive

The temptation is to go full agent on everything. Don't.

Anthropic's own recommendation — and my experience backs this up — is to start simple and add complexity only when it measurably improves the result. A single well-crafted prompt beats a complex agentic workflow for most tasks. Here's how I decide:

The progression matters. Each level adds latency, cost, and the chance of compounding errors. An agent that takes three wrong turns burns time and tokens. A well-scoped prompt that hits the target in one shot is always better — when the task allows it.

What Makes an Agent Work Well

After months of using Claude agentically across client engagements, here's what I've found actually matters:

Good tools beat good prompts. This is counterintuitive, but Anthropic found the same thing building their own agents: they spent more time optimizing the tools than the prompts. When Claude has well-connected, well-documented tools — a properly configured Google Drive, a structured Notion workspace, a clean Figma file — the agentic behavior is dramatically better. Garbage tools produce garbage agent loops.

Context is everything. An agent without memory, project instructions, and uploaded reference documents is just a general-purpose AI clicking through your tools randomly. The entire customization stack we've covered in this series — memory, projects, preferences — is what makes an agent effective. It's the difference between handing a task to a stranger and handing it to someone who's worked with you for six months.

Human checkpoints matter. I don't let Claude send emails, publish content, or modify client documents without reviewing the output. The agent does 90% of the work. The last 10% — the judgment call — stays with me. This isn't a limitation. It's the design.

Simplicity wins. The most complex system I've seen fail was a multi-agent pipeline someone built with four chained models and a custom orchestration layer. The most effective agentic behavior I use daily is a single Claude conversation with memory, project context, and connected tools. No framework. No abstraction layer. Just a well-configured AI with access to where my work lives.

Where This Is Going

I'll be honest: we're early. The agentic capabilities in Claude today are genuinely useful for my work, but they're not autonomous. Claude doesn't run my consulting business while I sleep. It doesn't replace the judgment calls, the relationship management, or the strategic intuition that clients actually pay for.

What it does is eliminate the grind. The research phase. The context-gathering. The first-draft assembly. The cross-referencing across tools. The formatting and document creation. The repetitive tasks that used to eat two hours of every day and required zero expertise — just patience.

That's what agents mean in practice right now. Not artificial general intelligence. Not autonomous businesses. Just a well-configured AI that can string together multiple steps, use your actual tools, and deliver a result that's 90% of the way there.

The remaining 10% is your job. And honestly, that's the part you should be spending your time on anyway.