How to build an AI agent without code

The job, stated so it can be checked

The agent takes a topic and returns a brief I can make a decision from: what the options are, what each costs, what the trade-offs are, and what I should do next. Not an essay.

Done-condition: the brief names at least three specific options, each with a price or a plainly stated 'no public price', a one-line trade-off, and a linked source. If any of those is missing, it is not finished.

Step 1 — Pick the surface

Any assistant builder that lets you set persistent instructions and switch on web search will do. What matters is that three things exist: instructions that persist between runs, at least one tool, and somewhere the output lands. If your builder cannot do all three, it is a prompt box, not an agent host.

• Persistent instructions — so you are not re-explaining the job every time.
• Web search or file access — the tool it actually needs.
• A destination — a doc, a sheet, or an email draft. Output with nowhere to go gets lost.

Step 2 — Write the instructions

Instructions are a job description, not a prompt. Role, standards, format, limits. Keep them short enough that you would read them yourself.

ROLE
You are a research assistant producing decision briefs for a one-person business owner. Your reader is short on time and will act on what you write.

HOW YOU WORK
1. Restate the question in one line so a misunderstanding surfaces immediately.
2. Search for current sources. Prefer primary sources: the vendor's own pricing page, official documentation, the original announcement.
3. Identify at least three genuine options. If there are fewer than three, say so and explain why.
4. For each option record: what it is, the price (or 'no public price'), who it suits, one honest trade-off, and the source link.
5. Finish with a recommendation and the single thing that would change your mind.

STANDARDS
- Never state a number you did not read in a source. If a figure is unavailable, write 'not published'.
- Do not describe marketing copy as evidence.
- If sources disagree, show both and say which you trust and why.
- Plain language. No preamble, no summary of what you are about to do.

FORMAT
A table of options, then a short recommendation, then 'What would change my mind'.

DONE WHEN
Every option has a price field, a trade-off, and a working source link.

Step 3 — Give it the smallest set of tools

Web search, and nothing else, for this one. Every extra tool is another decision the agent has to make correctly, and the first version of any agent is worse with more choices. Add tools when a run fails for want of one — never in advance.

Step 4 — Run it on something you already know the answer to

This is the step people skip, and it is the only real test. Run it on a topic you have already researched by hand. You will immediately see whether it found what you found, missed the obvious option, or invented a price.

I keep three known-answer topics as a standing test set. Any time I change the instructions, I re-run all three before trusting it on something new.

Step 5 — Fix the failures, not the model

Nearly every bad output traces back to the instructions, not the model. Three patterns cover most of it.

• Vague output → the format section is too loose. Name the exact fields.
• Invented figures → add the 'never state a number you did not read' line and require a source link per row.
• Stops too early → the done-condition is not checkable. Rewrite it as a list of fields that must be non-empty.
• Wanders and burns cost → too many tools, or the job is really two jobs. Split it.

Step 6 — Give the output somewhere to live

An agent whose output lands in a chat window gets used twice and abandoned. Point it at a document, a sheet row, or an email draft addressed to you. The moment the result shows up where the work happens, you start using it without deciding to.

What breaks in week two

The novelty wears off and you notice the brief is competent but generic. That is a scope problem: the agent does not know your situation. The fix is a short 'context' block in the instructions — who you are, what you sell, what you have already ruled out. Two paragraphs of that changes the output more than any model upgrade.

The other week-two failure is silent drift. Sources go stale, a pricing page changes shape, and the agent quietly starts guessing. That is why the source link per row is not decoration — it is how you catch it.