How to Build an AI Agent Without Writing Code
Build a working AI agent with no code: describe the task, upload documents, pick models, connect tools. What agents can and cannot do, and what it costs in NPR.
“Build an AI agent” sounds like something that requires a framework, a vector database and a weekend. For most useful agents, it does not. What it requires is a clear description of the job and the documents the agent needs to read.
Here is what actually goes into one, and where the real limits are.
What an agent is, minus the hype
An agent is a chat model with three additions:
- Memory that survives between sessions, so you stop re-explaining context.
- Retrieval over documents you supply, so answers are grounded in your material rather than the model’s general impressions.
- Tools it can call, so it can do a thing rather than describe the thing.
Take any one away and you have something less useful. Take all three away and you have a chat window.
Step 1: Describe the job in plain language
The single biggest predictor of whether an agent works is whether you can state what it should do in a couple of sentences. Vague in, vague out.
Weak: “help me with my documents.” Better: “read the invoices I upload, pull out the amount and due date, and tell me which are overdue as of today.”
The second one is testable. You can look at the output and say whether it is right, which means you can improve it.
Step 2: Give it something to read
Upload the PDFs, spreadsheets or notes it should be grounded in. This is the step that converts a generic model into something that knows your situation.
Two things worth knowing:
- Quality beats quantity. Ten relevant documents outperform two hundred where most are noise.
- Scanned images are not text. A PDF that is a photo of a page needs OCR first, or the agent sees nothing. Devanagari text is handled fine when it is real text.
Step 3: Pick which models it may use
Different models are genuinely better at different things, and they cost differently. A sensible agent uses a cheap fast model for mechanical extraction and a stronger one for judgement calls.
If you are unsure where to start, our GPT vs Claude vs Gemini comparison covers which tends to win at what.
Step 4: Connect tools, only if you need them
This is the step most guides overweight. Tool access via MCP matters when the agent must reach a live system — a database, an internal API, a file store. If everything it needs is in the documents you uploaded, skip it.
Step 5: Test it against cases you already know the answer to
Do not evaluate an agent on questions you cannot check. Feed it five documents whose correct answers you already know. If it gets those right, extend it. If it does not, the fix is almost always clearer instructions or better documents, not a different model.
Where agents genuinely fall short
Worth being blunt, because overselling this helps nobody:
- They are confidently wrong sometimes. Grounding in documents reduces this substantially; it does not eliminate it.
- They should not take irreversible actions unsupervised. Drafting the email is fine. Sending it to a customer list unreviewed is not.
- They know nothing you did not give them. An agent with no documents and no tools is just a chat model with a system prompt.
The useful mental model: an agent is a fast, tireless junior who has read everything you gave it and nothing you did not, and who will never tell you it is unsure unless you ask.
What it costs
There is no subscription on Nuvaa. You start free at Rs 0, top up credits from Rs 100, and the agent spends credits while it runs and not otherwise. An agent triggered a few times a week costs very little; one processing hundreds of documents daily costs more. You are not paying a flat monthly fee for a month you barely used it.
Ready to try? Build an agent free — no code, no card, and it reads Nepali documents as well as English. Or start with what an AI agent actually is if you want the concepts first.