You can build an AI agent yourself with Claude Code or Cursor when it is internal, low-stakes, reads rather than acts, touches one or two systems, and someone will maintain it. AI coding tools make the code cheap. What they do not supply is a written list of what the agent must never do, a test set built from your real cases, a security model against prompt injection, and someone measuring it after model updates. Customer-facing agents, agents that send, pay or change records, and agents handling sensitive data need those, which is where a professional build's cost goes.
When can you build an AI agent yourself with Claude Code or Cursor?
More often than an AI development firm would usually admit. Claude Code, Cursor and tools like them have made the first working version of most AI features cheap. Someone who can read code and run a terminal can get an agent answering questions over their own documents, or drafting replies from a CRM record, in a day or two.
Building it yourself is usually the right call when all of these are true:
- The users are your own team, not your customers.
- A wrong answer is an inconvenience, not a refund, a compliance issue or an angry client.
- The agent reads and suggests. It does not send, pay, delete or update anything on its own.
- It touches one or two systems you already have API access to.
- Someone on your side will still be around to fix it in six months.
An internal search tool over your policies, a script that drafts first replies for a person to edit, a prototype to show investors or test with five friendly users: build these yourself. Paying a firm $4,500-8,000 for something you could build in a weekend and nobody outside the company sees is poor value, and we say so on calls.
What are AI coding tools genuinely good at?
We use them every day, so this is not a dismissal. They are very good at the parts of an AI build that are ordinary software:
- Wiring up a model provider's SDK, streaming responses and handling retries.
- Writing the integration code for a well-documented API.
- Building a chunking and embedding pipeline for a vector database.
- Producing a working interface - a chat widget, an admin screen, a Slack bot.
- Writing unit tests for code whose correct behaviour you can describe.
- Explaining an unfamiliar codebase or library faster than reading the docs.
That covers a large share of the hours in a typical build. The honest consequence is that AI development should be getting cheaper, and quotes that have not moved since 2024 deserve a question.
What do AI coding tools miss when you build an agent?
The tools write the code you ask for. The expensive failures in AI systems are the ones nobody thought to ask about, and the tool cannot see those either. Five come up again and again.
1. Nobody decided what it must never do
An agent with write access to a CRM, an inbox or a payments system needs a written boundary: what it may do on its own, what needs approval and what it may never do. A coding tool will happily build an agent that can send emails. It will not stop and ask whether it should send them without a person seeing them first.
2. There is no test set from real cases
Unit tests check that code does what you described. They cannot tell you whether the agent gives the right answer to the forty awkward questions your customers actually ask. That needs a labelled evaluation set built from your own tickets, documents or leads, and a score agreed before launch. It is tedious, it is the most valuable part of the build, and it is the part DIY projects skip. Our post on evaluating LLM output in production explains how to build one.
3. It works on the demo and fails on the edge
The prototype answers the five questions you tried. In production it meets scanned PDFs, tables, product codes, questions in two languages and documents that contradict each other. Retrieval systems in particular fail quietly - they return a confident answer from the wrong document. We wrote up the warning signs in 7 signs your RAG chatbot is giving wrong answers.
4. Security is someone else's problem
Any agent that reads text from outside - emails, web pages, uploaded files - can be instructed by that text. Prompt injection is not solved by a better prompt. It is solved by limiting what the agent can do, separating its permissions per user and keeping secrets out of its reach. A coding tool will put the API key where you tell it to.
5. Nobody owns month six
Model providers retire and update models on a published schedule. Documents change. Usage grows and the token bill grows with it. A DIY agent tends to degrade silently because nobody is measuring it, and the first sign is a customer complaint.
How does building it yourself compare with hiring a firm?
Side by side, the difference is less about the code and more about everything around it.
| Build it yourself | Hire a firm | |
|---|---|---|
| Cost to first version | Your time plus model costs | From $4,500 fixed |
| Time to first version | Days | 2-4 weeks for a single-purpose agent |
| Evaluation set and threshold | Only if you build one | Agreed before build |
| Approval paths and permissions | Only if you design them | Written into the scope |
| Security review | Your judgement | Part of the build |
| Maintenance after launch | Whoever wrote it | Handover, runbooks, optional support |
| Best for | Internal tools, prototypes, low stakes | Customer-facing or action-taking systems |
How do you decide whether to build it yourself?
Answer these honestly. Each "yes" is a reason to bring in someone who has shipped this before, at least for a review.
- Will customers, not just staff, see the agent's output?
- Can the agent send, pay, book, delete or change records without a person approving?
- Does it handle personal, financial, health or legal data?
- Does it need more than two integrations?
- Would a wrong answer cost more than a day of someone's time to fix?
- Is there nobody on your side who will maintain it next year?
No to all six: build it yourself, and spend the money you saved on a good evaluation set. One or two: build it, then pay for a review before it goes live. Three or more: the parts you would skip are the parts that matter.
Is there a middle route between DIY and hiring a firm?
Yes, and it is often the best value. Build the prototype yourself with AI coding tools, prove that the idea works and that people use it, then bring in engineers to harden it.
- 01
Prototype it yourself
Get a working version in front of real users quickly. The goal is to learn whether it is worth building properly, not to ship it.
- 02
Collect the hard cases
Every question it got wrong, every document it misread. This becomes the evaluation set, and it is worth more than the prototype code.
- 03
Pay for a review
An experienced engineer reads the code, the prompts and the permissions, and lists what would break in production. Often a few days of work, not a rebuild.
- 04
Harden or rebuild
Sometimes the prototype becomes the product with fixes. Sometimes it is cheaper to rebuild it properly using what you learned. Either way, you know.
If you are an agency weighing this for a client project, the same logic applies, and our guide to delivering AI without hiring covers the options.
Can you build an AI agent yourself with Claude Code or Cursor?
Yes, if it is internal, low-stakes, reads rather than acts, and someone will maintain it. AI coding tools have made the code the cheap part. What they do not supply is the written boundary of what the agent must never do, a test set built from your real cases, a security model and someone watching it after launch - and those are what separate a demo from a system customers can rely on.
If your answers to the checklist point the other way, see how we price on why we're not the cheapest, or AI agent development for what a professional build includes.