What are Claude agents? A plain-English guide | satori.

Heard the words "AI agent" more often this past month than in the whole of last year? You're not alone. Claude agents are one of the fastest-growing Claude searches in Australia right now¹, and it's no accident. In April, Anthropic opened its agent platform, Claude Managed Agents, to every developer². And since June, agents can be scheduled, so they work entirely on their own while you do something else³.

But what actually is an agent? Is it just a fancier chatbot? And what do all the flying words mean: managed agents, subagents, Agent SDK?

Let's sort the whole picture out, using Anthropic's own definitions as the map. So you don't have to guess.

---

What are Claude agents, really?

Start with the basic question. A regular AI chat is a sounding board: you ask, it answers, and then you do the work yourself. An agent takes the work.

Anthropic's own definition separates two things. Workflows are systems where the AI follows predefined code paths, a bit like a recipe. Agents are systems where the model directs its own process and its own tools, like a chef who opens the fridge and decides what's for dinner⁴. In practice, an agent is a model using tools in a loop: gather context, act, check the result, repeat. Until the task is actually done.

And this isn't theory anymore. Anthropic's own usage data shows automation overtook augmentation during 2025: the share of conversations where the AI performs tasks independently rose from 27 to 39 per cent, and for business customers via the API it sits at 77 per cent⁵. Meanwhile, the longest agent sessions in Claude Code have nearly doubled, from under 25 minutes to over 45⁶. The agents keep working longer and longer without anyone holding their hand.

But here's where the confusion starts. "Claude agents" is an umbrella over four different things: Managed Agents on the developer platform, subagents in Claude Code, the Agent SDK for building your own, and the desktop agent Claude Cowork. Let's take them one at a time.

---

What are Claude Managed Agents?

This is the news driving the search wave. On 8 April 2026, Anthropic launched Claude Managed Agents in public beta: a ready-built agent harness that runs on Anthropic's managed infrastructure².

Think of the difference between building your own restaurant kitchen and renting a fully fitted commercial one. Developers used to build the agent loop, the tool execution and the secure sandboxes themselves. Now all of that comes included: Claude can read files, run commands, browse the web and execute code in a secure environment that Anthropic operates⁷. Anthropic describes it as the path from prototype to launch in days instead of months, and companies like Notion, Rakuten, Asana and Sentry are already building on the platform².

Claude Managed Agents: the four building blocks

The platform rests on four concepts⁷. An agent is the configuration itself: model, instructions, tools. Created once, reused forever. An environment is the kitchen where the agent works, either Anthropic's cloud sandbox or your own infrastructure. A session is a shift. And events are the running commentary in and out, so you can follow exactly what the agent is doing.

Claude Managed Agents: scheduled agents

The June 2026 update is what makes this genuinely interesting for ordinary businesses: scheduled agents. Give the agent a cron schedule and it starts a fresh session every time the schedule fires, with no scheduler of your own to build or babysit³.

The Monday report done by 6 am. The week's supplier invoices checked before the finance team has had their first coffee. That's the kind of work agents are built for.

---

What do subagents do in Claude Code?

Subagents are something else entirely, and this is where most people mix up the cards. A subagent lives inside a Claude Code session. Anthropic defines them as specialised AI assistants for specific types of tasks, where each subagent runs in its own context window with its own system prompt, its own tools and its own permissions⁸.

Think of the main agent as a project manager. When a big research job turns up, the project manager delegates to a specialist. The specialist reads a thousand lines of code or thirty documents, and hands back a ten-line summary. All the noise stays with the specialist. The project manager's desk stays clean.

In practice, a subagent is just a markdown file in the .claude/agents/ folder, or something you create with the /agents command. And they have clear boundaries: a subagent can't start subagents of its own, and it works within a single session⁸. A smart detail for the wallet: you can route simpler subtasks to faster, cheaper models while the main agent runs the heavy one.

We use this ourselves every single day, including when we build agent workflows for clients. It's like having a small team instead of one lone generalist.

---

Managed Agents, Agent SDK or Claude Code: which one do you pick?

The third puzzle piece is the Claude Agent SDK, a library in Python and TypeScript that gives you the same tools, agent loop and context management that power Claude Code, but in your own code on your own infrastructure⁹.

Anthropic's own rule of thumb is blunt: prototype locally with the Agent SDK, move to Managed Agents when it's time for production⁹. Here's how it fits together:

And wherever the agent lives, it can load skills: folders of instructions, scripts and resources the agent picks up when the task calls for them¹⁰. The same skill works in the Claude apps, in Claude Code and on the platform. Write once, use everywhere.

For businesses without a developer team, the path is usually simpler than it sounds. An agent that takes your meeting transcripts, pulls out the decisions and updates your CRM through an MCP connection is a build of days, not months. That's exactly the kind of thing we set up in our build services.

---

What do Claude agents cost, and are they safe for company data?

The pricing is unusually simple. Managed Agents cost normal token prices plus US$0.08 per active session hour, roughly A$0.12, for the hosting itself². Subagents are included in your Claude Code usage, and the Agent SDK is billed as ordinary API tokens.

The privacy question deserves more honesty. Managed Agents are built to remember their sessions, and precisely for that reason the beta is not yet covered by Anthropic's zero data retention options⁷. Sessions and files can be deleted via the API whenever you like, but automatic zero storage isn't there yet. Our advice: run Managed Agents happily for processes without personal information, and keep the personal-data-heavy flows in Claude Code or the Agent SDK, where the data stays with you. Since processing happens in the US, your APP 8 cross-border homework applies here too. We've written a full guide to Claude and the Australian Privacy Act if you want the details.

This is exactly the kind of judgement call we help clients make in our managed Claude service: the right agent form for the right data, without compliance becoming a handbrake.

---

Summary: from tool to colleague

An AI agent is a model that uses tools in a loop until the job is done. Claude Managed Agents are Anthropic's hosted version of that, subagents are the specialists inside Claude Code, and the Agent SDK is the kit for those who want to own the whole stack. Different doors into the same future: AI as a workmate rather than an encyclopaedia.

Checklist: your first Claude agent

• Pick one well-bounded task that already happens every week, like a recurring report
• Decide what the agent may read and do before it gets to do anything at all
• Test with a human in the loop for the first few weeks and review the output
• Measure the time you save, or you'll never know whether the agent earns its keep
• Scale only after it's proven itself, one agent at a time

A tip from satori.: automation is step five on the AI ladder, not step one. Platform, training and structured data first. Then your first agent becomes a lever instead of a half-finished experiment.

Want help going from curious to an agent in production? Get in touch and we'll look at where an agent would do the most good for you.

---

Sources

Footnotes

1. Google Trends (2026). "claude agents", Australia, past 12 months. https://trends.google.com/trends/explore?geo=AU&date=today%2012-m&q=claude%20agents ↩

2. Anthropic (2026). Claude Managed Agents. https://claude.com/blog/claude-managed-agents ↩ ↩² ↩³ ↩⁴

3. Anthropic (2026). What's new in Claude Managed Agents. https://claude.com/blog/whats-new-in-claude-managed-agents ↩ ↩²

4. Anthropic (2024). Building effective agents. https://www.anthropic.com/engineering/building-effective-agents ↩

5. Anthropic (2025). Anthropic Economic Index: September 2025 report. https://www.anthropic.com/research/anthropic-economic-index-september-2025-report ↩

6. Anthropic (2026). Measuring agent autonomy. https://www.anthropic.com/research/measuring-agent-autonomy ↩

7. Anthropic (2026). Managed Agents overview, Claude Platform docs. https://platform.claude.com/docs/en/managed-agents/overview ↩ ↩² ↩³

8. Anthropic (2026). Subagents, Claude Code docs. https://code.claude.com/docs/en/sub-agents ↩ ↩²

9. Anthropic (2026). Agent SDK, Claude Code docs. https://code.claude.com/docs/en/agent-sdk ↩ ↩²

10. Anthropic (2025). Introducing Agent Skills. https://claude.com/blog/skills ↩