Claude agents for operations: what they actually do
An agent decides its own next step. Here is what that means for an operations team, which jobs suit it, and the four limits we use.
Short answer: an agent is a program where Claude decides its own next steps and uses tools to get a job done, such as reading a HubSpot deal, checking an invoice and drafting a reply. A workflow follows steps you wrote in advance. A chat waits for a person to ask. Most business tasks need a workflow, and a few need an agent.
Below is how the three differ, what the Claude Agent SDK is, which operations jobs suit an agent, and the four limits we put around every one.
Agent, workflow or chat?
Anthropic’s engineering team draws the line this way: workflows are systems where models and tools are orchestrated through predefined code paths, and agents are systems where the model directs its own process and tool use. The same article recommends finding the simplest solution possible and adding complexity only when needed.
| Who decides the steps | Good for | |
|---|---|---|
| Chat | A person, one message at a time | Drafting, questions, one-off analysis |
| Workflow | Your code or HubSpot, in a fixed order | Tasks with the same steps every time |
| Agent | Claude, within the tools and limits you give it | Tasks where the steps depend on what it finds |
Anthropic notes that agents trade higher cost and delay for flexibility, and that mistakes can add up over many steps. So if you can write the steps down, write them down. Our post on AI agents for business goes through that choice in more detail.
What the Claude Agent SDK is
The Claude Agent SDK is a code library, available in Python and TypeScript, that gives a developer the same agent loop, built-in tools and context management that power Claude Code. According to Anthropic’s documentation it includes permissions (which tools run automatically and which need approval), hooks (custom code that runs at set points in the agent’s work), subagents, sessions, and MCP, the standard way to connect external tools and data. It was previously packaged as the Claude Code SDK, and the docs carry a migration guide.
It is not a product you log in to. Someone has to build the agent with it, and run it in a process you control. Anthropic also offers Managed Agents, where Anthropic hosts the agent, as a separate option. For most mid-market businesses the choice between the two is a build decision for your engineer, not something your team will see.
Jobs that suit an agent
- Supplier invoice checking. The agent reads an invoice, looks up the purchase order, finds the mismatch, and prepares a note for accounts to approve.
- Deal desk preparation. Before a renewal call, it pulls the HubSpot deal, recent tickets and emails, and writes a one-page brief.
- Onboarding follow-up. It checks which steps a new client has finished, flags what is stuck and drafts the nudge for the account manager.
- CRM data repair. It investigates why a set of contacts has no owner, proposes fixes, and applies them after approval.
- Quote chasing. It finds quotes with no reply after a set time and drafts follow-ups in the voice of the owner.
Each one involves looking something up, deciding what to look at next, and producing something a person can check. Jobs that never vary, such as sending a welcome email when a form is submitted, belong in a HubSpot workflow.
The four limits
- Human approval for anything that leaves the building. The agent drafts emails, invoices and record changes. A person sends or approves them. The Agent SDK’s permission settings control which tools can run automatically and which need approval.
- Read before write. Start the agent with read-only access. Give it write access one field or one object at a time, once you have seen it behave.
- Logging. Every step, tool call and result is recorded, so when something looks wrong you can see what the agent saw and did. The SDK’s hooks are one place to capture this.
- A way to undo. Before an agent changes a record, save the old value. If a run goes wrong, you restore it rather than reconstruct it.
Thinking about an agent for your operations team?
Tell us the job and we will say whether it needs an agent at all.
We wrote up how we handle each in more detail. How we test an AI agent before it goes near a customer covers testing, and can you undo what an AI agent just did covers rollback. Our HubSpot and MCP post shows read-before-write in a real build.
How to start small
Pick one job that a person does weekly and finds tedious, where a wrong answer is easy to catch. Run the agent in read-only mode for two weeks and compare its output with what your team would have done. Fix the instructions where it differs. Then let it prepare drafts, and only after that let it make small, reversible changes.
Anthropic’s own advice is to test agents extensively in sandboxed environments and to set stopping conditions, such as a maximum number of steps. We add a budget cap and a daily summary of what the agent did.
We’re Certified in the Claude Partner Network, and our AI engineering team builds agents like these for Australian businesses on top of Claude. See our Claude for business page for how we work, and the Claude pricing guide for what the model usage costs.
Frequently asked questions
What is a Claude agent? It is a program where Claude plans its own steps and calls tools, such as reading a CRM record or an invoice, to finish a task. A person sets its limits and approves what matters.
What is the Claude Agent SDK? It is Anthropic’s Python and TypeScript library for building agents with the same loop, tools and permissions that power Claude Code. A developer uses it to build the agent.
What is the difference between an agent and a workflow? A workflow follows steps written in advance. An agent decides its next step from what it finds along the way.
Is agentic AI safe for business use? It can be, with human approval, read-only access to start, full logging and a way to undo changes. The risk comes from giving an agent write access before you have watched it work.
Where should a business start with agents? Start with one weekly task where mistakes are easy to spot, and run the agent read-only first.
Sources
- Agent SDK overview, Claude documentation, checked 31 October 2026, for what the SDK includes, its languages and the Claude Code SDK migration.
- Building effective agents, Anthropic engineering, checked 31 October 2026, for the workflow and agent definitions, and guidance on simplicity, checkpoints and testing.
Ready to try an agent on one real job?
A narrow first build teaches you more than a broad plan.