The next level beyond workflow automation is AI agents — systems that can understand a goal, break it into steps, execute those steps, observe the results, and adapt. Here’s how we build them.
What Makes Something an “AI Agent” vs. an Automation?
A regular automation follows a fixed path: if X, then Y, then Z. An AI agent can reason about what to do next based on intermediate results.
Automation example: New form submission → send SMS → wait 24 hours → send email
AI Agent example: New form submission → Claude reads it → decides which follow-up strategy makes most sense → executes that strategy → reads the response → decides next step accordingly
The Architecture
Our multi-step agents typically look like this in n8n:
- Input node: Receives the trigger (webhook, schedule, API call)
- Context node: Pulls relevant data from CRM, email history, previous interactions
- Claude reasoning node: Feeds all context to Claude with instructions and a list of available tools
- Tool execution nodes: Send SMS, update CRM, create calendar event, etc.
- Observation loop: Claude reviews what happened and decides whether to continue or stop
Real Example: Dormant Lead Reactivation Agent
We built an agent that runs weekly, identifies leads that went cold 30-90 days ago, researches each one (checks if their website changed, if they have new Google reviews, if their business has grown), then crafts a highly personalized reactivation message for each one.
One client recovered $47,000 in deals in the first month from dormant leads. Ask us about building an agent for your business.
Guardrails We Build Into Every Agent
Limiting Which Tools the Agent Can Call
An agent should only be able to do what its job requires. A reactivation agent might read the CRM and send texts, but it has no reason to delete contacts or issue refunds. We expose a short, explicit tool list to Claude, which keeps behavior predictable and makes it far easier to reason about what could go wrong.
Capping Loops and API Spend
Reasoning loops can repeat if a tool keeps returning an unexpected result. Every agent we build has a maximum number of iterations per run and a daily usage ceiling. When either limit is hit, the run stops and alerts a person instead of quietly burning through API credits overnight.
Requiring Approval for High-Stakes Actions
Some steps deserve a human glance: messages to large accounts, anything involving money, or replies to an upset customer. For those, the agent prepares the action and posts it for one-click approval rather than executing it directly. Routine actions run on their own; sensitive ones wait for a yes.
Logging Every Decision for Review
Each run records what the agent saw, what it decided and why, and what each tool returned. When a result looks odd, we can trace exactly where the reasoning went off track and adjust the instructions. Those logs also become training examples for improving the prompt over time. Our guide to Claude explains why its reasoning suits this work.
Starting With a Narrow Job
The agents that succeed do one well-defined job before they are given another. Begin with something contained, like dormant lead reactivation or appointment confirmation, prove it works, and expand from there. Our n8n and Anthropic automation projects usually start this way, often connected to GoHighLevel as the system of record.