Tutorial · 2026
Build your first AI agent
In one line: Scope one job, baseline the metric, ground the agent in your own documents, add a human confirmation step — then run it for two weeks and read the transcripts.
What you are actually assembling
An agent is not a bigger chatbot. It is four parts working together: perception (it can read text, files, web pages), memory (short-term context plus a long-term store), planning (breaking a goal into steps) and action (calling tools to change something in the world). If any one of those is missing, you have a chatbot with ambitions.
Six steps from zero to a working pilot
- Write down one job in one sentence. “Answer every WhatsApp enquiry about opening hours, pricing and booking within 30 seconds, and hand anything else to staff.” If you cannot state it in one sentence, the scope is too big.
- Define what success looks like before you start. First response time, percentage resolved without a human, hours saved per week. Without a number, every result is arguable.
- Give it real knowledge. Load your price list, FAQ, policies and product documents. An agent grounded on your documents is dramatically better than one improvising from general training data.
- Give it the smallest set of tools that does the job. Reading the knowledge base and sending a WhatsApp reply may be enough. Every extra tool is extra risk.
- Add a confirmation step for anything irreversible. Payments, refunds, commitments and deletions should require a human yes.
- Run it for two weeks, then read the transcripts. The failures tell you what to fix: missing knowledge, ambiguous policy, a missing tool, or a badly written instruction.
The mistakes that stall most projects
- Starting with technology. Teams pick a framework first and look for a use case second. Reverse it.
- No knowledge base. The agent answers plausibly but wrongly, staff lose trust, and the project is quietly abandoned.
- No escalation path. One bad customer interaction poisons the whole rollout. Always define what gets handed to a human.
- Measuring nothing. A pilot with no baseline cannot prove value and therefore never gets budget.
- Building instead of buying when there is no engineering team. A framework without engineers is a hobby, not a product.
What it costs to start
With a platform such as MOKAGI, the entry point is a one-off HK$5,000 installation including agent configuration and the first knowledge base, usage at HK$68 per million tokens, and HK$299 per month upwards for an AI customer-service agent. For a single pilot scenario, that is small enough to approve without a committee.
A realistic timeline
| Week | What happens |
|---|
| 1 | Scope, baseline metrics, knowledge collection |
| 2 | Install, configure the agent, connect the first channel |
| 3–4 | Supervised running, transcript review, fixes |
| 5 | Decide: scale to the second scenario, or stop |
Frequently asked questions (FAQ)
How long does it take to build a first AI agent?
A focused pilot usually takes two to four weeks: one week to scope and prepare knowledge, one to configure, and two to run under supervision.
Do I need to write code?
Not if you use an agent operating system. Frameworks such as LangChain do require development work.
What should the first use case be?
Something repetitive, high volume and low risk — usually customer-service FAQs or internal knowledge Q&A.
How do I know it is working?
Compare against a baseline recorded before launch: first response time, resolution rate without a human, and hours saved per week.