How to build an AI agent: no-code, framework or custom?

Can you build an AI agent yourself?

Often, yes. A capable person can configure a useful internal agent in a no-code platform, and a developer can code one with a framework. The first version comes quickly. The hard part comes after: making the agent right often enough, on real inputs, with permissions you can defend, and keeping it that way as your systems change.

This guide is written for the person who has to decide how to proceed. It covers the three routes, what each one can and cannot do, the steps that matter whichever you choose, and the cases where an agency is worth paying for. We build agents for a living, so we say plainly where you do not need us.

One caution before starting. Anthropic, whose engineering guide on agents dates from December 2024, now opens it with a note that “much of the tooling landscape described in this post has changed since December 2024”. Expect the tools below to keep moving; the method in this guide has held up through those changes.

First, check that the task needs an agent

Build an agent only when the work needs judgment on variable input; otherwise a plain workflow is cheaper and safer. OpenAI’s guide to building agents lists three signs that an agent is the right tool: complex decisions with exceptions, rules that have become too hard to maintain, and heavy reliance on unstructured data such as emails and documents. If your case does not clearly meet them, it says, “a deterministic solution may suffice”.

Anthropic makes the same point in stronger terms: start with “the simplest solution possible”, which “might mean not building agentic systems at all”. Gartner, which predicts that over 40% of agentic AI projects will be canceled by the end of 2027, adds that “many use cases positioned as agentic today don’t require agentic implementations”.

A quick test: can you write the steps on a single page, and are they the same every time? Then you need a workflow, possibly with one AI step to read a document or classify a message. If the steps depend on what the case turns out to be, an agent makes sense. Our guide What is an AI agent? explains the difference in more detail, and our guide to agentic AI shows the range of options between the two.

Three ways to build an AI agent

The three routes differ in who builds, how much you control, and what you own at the end.

These routes are not exclusive. A common path is to prove the value on a platform, then rebuild the part that works on a framework when volume, integration or compliance demands it.

Route 1: no-code and low-code platforms

Platforms let you build a working agent without writing code, as long as it lives inside the tools they connect to. They are the right place to start for an internal helper or a proof of value.

Three examples show the range:

• n8n is a workflow automation tool with an AI Agent node. Its documentation describes it plainly: “Connect a chat model and one or more tools, and the agent decides which tools to call to complete a task.” n8n can be self-hosted on your own infrastructure, and without a license key it runs as the free Community edition. That makes it a common choice for companies that want to keep data on their own servers. Our guide to the n8n AI agent covers the nodes, tools and human approval in detail, and our n8n self-hosted guide covers the license and the setup.
• Microsoft Copilot Studio is, in Microsoft’s words, “a graphical, low-code studio for building and managing AI-powered agents and workflows”, connected to your organization’s data and published in Teams, Microsoft 365 Copilot or a website. It fits companies already on Microsoft 365.
• OpenAI Agent Builder, part of the AgentKit release of October 6, 2025, is “a visual canvas for creating and versioning multi-agent workflows”, with guardrails and evaluations configured in the same place.

What you can honestly do alone on these platforms: triage and route emails, draft replies for approval, summarize tickets, fill a spreadsheet or CRM field from a document, answer internal questions from a set of files.

Where they reach their limits: systems without a ready-made connector, business rules too intricate to express visually, testing on hundreds of real cases before each change, fine-grained permissions per action, and traceability an auditor will accept. You also depend on the platform’s pricing, its roadmap and its choice of models.

Route 2: agent frameworks

A framework gives developers the agent loop, tool calling and tracing, so they write only the business logic. It is the route for teams that code and want control without starting from nothing.

The differences matter less than they seem. Pick the one that matches your language, your model provider and your hosting, and make sure you can see what happens at each step.

Anthropic’s warning applies to all of them: frameworks “often create extra layers of abstraction that can obscure the underlying prompts and responses, making them harder to debug”. Its advice is to start with the model’s API directly where possible and, if you use a framework, to “ensure you understand the underlying code”. Frameworks “can help you get started quickly”, it concludes, “but don’t hesitate to reduce abstraction layers” in production.

Route 3: a custom build inside your systems

A custom build is worth it when the agent must act inside the systems that run the company and prove its reliability there. The agent itself is often the smaller part of the work. Connecting it to an ERP, a CRM or an old in-house database, with the right rights and a full log, is the larger one.

Gartner warns that “integrating agents into legacy systems can be technically complex, often disrupting workflows and requiring costly modifications”, and suggests that rethinking the workflow around the agent is often better than bolting the agent on.

The signals that point to a custom build:

• The agent writes into core systems as well as reading them.
• It handles personal data, contracts or payments, so GDPR and possibly the AI Act apply.
• Your data or models must stay in Europe, or on your own servers.
• Volume is high enough that the cost per task and the error rate have to be measured.
• It is part of a product you sell, so you need to own the code.

If several of these apply, see how we build AI agents for business and how we connect them to existing tools through AI integration services. For a first idea of budget, our guide to AI agent cost gathers published market ranges.

The steps that matter, whichever route you choose

The same seven steps apply on a platform, on a framework or in a custom build. Skipping one is the usual reason an agent never leaves the pilot stage.

1. Define one task and its indicator. Write what the agent receives, what it must produce, and how you will measure success: time saved, share of cases handled without correction, errors caught.
2. Collect real cases. Take past emails, tickets or documents, including the awkward ones, and write the expected outcome for each. This set becomes your test.
3. List the tools, with the fewest rights that work. OWASP recommends limiting the tools an agent may call “to only the minimum necessary”, and their permissions to the minimum too. Read-only first.
4. Write the instructions. State the rules, the tone, the limits and when to stop and ask. Document each tool as carefully as a screen for a human user; Anthropic suggests investing “just as much effort” in these agent-computer interfaces.
5. Set a baseline, then optimize. OpenAI recommends building the prototype with the most capable model to establish a performance baseline, then trying smaller models to see if they hold up.
6. Put a person on the sensitive actions. OpenAI names two triggers for human intervention: repeated failures, and actions that are “sensitive, irreversible, or have high stakes”, such as refunds or payments.
7. Test before release, and watch after. Run the test set on every change, in a sandbox first, as Anthropic advises. In production, log every action and review the failures each week.

You can try what a well-scoped result feels like on our document extraction demo: a fixed task, structured output, and checks that catch errors.

When to call an agency, and when not to

Build it yourself when the task is internal, the stakes are low, the tools are already connected and someone in-house can own it. That covers many first agents, and it is the fastest way to learn what works in your company.

Bring in an agency when one or more of these is true:

• The agent must act in core systems, with integration work and security reviews.
• Personal or regulated data is involved, and you need a design you can defend under GDPR and the AI Act.
• You need proof of reliability on your own cases before rollout, and a method to keep it measured.
• Nobody in-house has time to maintain the agent when models, APIs or your processes change.

There is a middle path. If you want your team to build and run its own agents, our AI training for employees works on your tools and your cases, and each team leaves with a first agent. If you are not sure where to start, an AI readiness assessment ranks the tasks where an agent will pay off first.

Whichever route you lean toward, a second opinion before you build costs nothing. Book your free 30-minute assessment: describe the task and the tools involved, and we tell you which of the three routes fits, what you can safely do in-house and where the risks sit. You leave with an opinion even if we never work together, and we reply within one business day.