Step 1: Define what your agent should do
Before writing a line of code or setting up a no-code builder, get precise about what the agent should accomplish. Vague goals produce vague agents.
A good agent definition answers three questions:
- What is the trigger? When does the agent run? (On schedule, on an event, on demand)
- What does it read? Which tools and data sources does it need to access?
- What does it write? What changes does it make, and where?
Example definition
Trigger: Every Monday at 8am
Reads: HubSpot: all open deals, last activity date, deal owner
Writes: Slack: posts a summary of stale deals (no activity in 7+ days) to #sales-ops, tagged by owner
Start with one recurring task. The agents that fail are almost always the ones that try to do too much in the first version.
Step 2: Choose your build approach
There are three ways to build an AI agent. The right choice depends on your technical skills and how much customization you need:
| Approach | Setup time | Requires code | Controls | Best for |
|---|---|---|---|---|
| No-code builder (Pinksheep) | Fastest to start | No | Built-in | Ops teams, non-engineers |
| Code-first framework (LangGraph, AutoGen) | Longer setup | Yes | Build it yourself | Engineers, custom logic |
| Hybrid (no-code + connected tools) | Varies | Minimal | Built-in | Teams needing custom data access |
Step 3: Connect your tools and data sources
An agent is only as useful as the tools it can access. Connections break down into two types:
- Native integrations. Pre-built OAuth connections to common tools such as HubSpot, Salesforce, Slack, Google Sheets, Jira, Notion, Zendesk, and Stripe. These reduce setup effort compared with building the connection yourself.
- MCP servers. Model Context Protocol connections for AI-native tool access. MCP-compatible tools can be connected to your agent where that integration path is available.
The principle of least privilege applies here: only connect the tools the agent actually needs. If your CRM cleanup agent doesn't need access to billing tools, don't connect them.
Step 4: Add guardrails
Controls are what separate production-ready agents from demo agents. There are four you should have before deploying an agent on any system that matters:
Approval gates
Every write action routes to a human for sign-off before it executes. The agent proposes; you confirm. This is the most important guardrail.
Audit trail
Every action the agent takes is logged: what it read, what it proposed, who approved it, and what executed.
Spend caps
Set a maximum spend limit per agent. When the agent hits the cap, it stops cleanly and notifies you.
Access scoping
Restrict the agent to only the tools and data it needs. A reporting agent should not have write access to your payment processor.
Pinksheep builds these controls into every agent by default. For code-first builds, you need to implement each one. See the human-in-the-loop AI agents guide for implementation details.
Also see: AI agent approval workflows, AI agent spend control.
Step 5: Test and deploy
Before scheduling an agent to run automatically, test it manually. The testing process:
- Dry run. Trigger the agent manually on a small, real dataset. Review every action it proposes before approving any of them.
- Verify reads. Confirm the agent is reading the correct data from the correct sources. Check for scope issues (reading too much or too little).
- Approve one action. Approve a single write action and verify it executes correctly in the target system. Check the action history entry.
- Run a full cycle. Let the agent run through a complete task with all approval gates active. Review the action history.
- Schedule deployment. Only once you have confirmed the agent works correctly on real data with approvals active, schedule it for automated recurring runs.
Build an AI agent without code
If you are using Pinksheep, here is the specific no-code path from idea to deployed agent:
- Start in Pinksheep. Open the builder and begin with a single recurring task.
- Connect your tools. Connect the systems the agent needs to read from or write to.
- Describe the agent. Type what you want in plain English: "Every Friday at 4pm, check all open HubSpot deals with no activity in 7 days and post a summary grouped by owner to #sales-ops."
- Review the plan. Pinksheep generates a step-by-step execution plan. Read every step. Adjust any that look wrong.
- Approve and deploy. Confirm the plan. The agent activates with approvals and visibility in place by default.
Build AI agents for your business
No code. No complexity. Just describe what you need.
For a full guide to the no-code approach, see no-code AI agents.
AI agent frameworks compared
If you are taking the code-first approach, here are the main frameworks and when to use each:
LangGraph
Best for: Complex multi-step workflows with conditional branching
Graph-based orchestration. Best for workflows where the path depends on what the agent finds. High learning curve.
AutoGen
Best for: Multi-agent systems where agents collaborate
Conversation-based multi-agent coordination. Good for tasks that benefit from agent specialisation and debate.
CrewAI
Best for: Role-based agent teams
Assigns roles (researcher, writer, reviewer) to agents in a crew. Simpler than AutoGen for structured agent teams.
smolagents (Hugging Face)
Best for: Lightweight agents with custom model choices
Minimal framework. Good for teams that want to use non-OpenAI models or need a very thin abstraction layer.
All code-first frameworks require you to build the control layer yourself, including approvals, visibility, and spend controls. If that is a concern, use a platform like Pinksheep or add that layer on top of your framework. See the best AI agent builders comparison for more options.
Common questions
How long does it take to build an AI agent?
That depends on the build approach, the number of systems involved, and how much review logic you need. A no-code builder can shorten setup significantly, while code-first builds usually take longer because you are assembling more of the control layer yourself.
What tools do I need to connect to my AI agent?
This depends on your workflow. Common integrations include CRM tools, communication tools, project management tools, databases, spreadsheets, and finance systems. Start with only the systems the agent actually needs.
Do I need to be a developer to build an AI agent?
No. No-code agent builders like Pinksheep let anyone describe an agent in plain English. The builder handles the technical execution. If you want maximum customisation, code-first frameworks give engineers full control but require Python experience and significant build time.
How do I prevent an AI agent from making mistakes?
Use approval gates on all write actions. The agent reads and reasons automatically, but every proposed change to an external system (CRM update, email send, file creation) routes to a human for sign-off before it executes. Start with approvals on everything, then relax controls incrementally on action types you trust.
What is an AI agent framework?
An AI agent framework is a code library that provides the core abstractions for building agents: tool calling, memory, planning, and multi-agent coordination. The main frameworks are LangGraph (graph-based orchestration), AutoGen (multi-agent conversation), and CrewAI (role-based crews). These are code-first tools for engineers. No-code builders sit above the framework layer.
Can I build an AI agent for free?
Pricing depends on the product you choose and how much usage your agent sees. Check current pricing directly rather than relying on generic estimates.