I've lived through a lot of crypto narratives since 2011: ICOs, DeFi, NFTs, layer 2s, memecoins, Ordinals, GameFi, real-world assets, restaking. Each one came with a wave of projects that were real, and a much bigger wave that only borrowed the name.
AI agents are the latest narrative, and I'm seeing the same pattern. Some teams are building genuinely useful things. Others are attaching "AI" to a token and hoping nobody asks what the agent actually does.
I build AI agents, chatbots and LLM automation for clients, and I also run one on my own website. So here's my honest view of what's worth building for a Web3 team, and how to do it without creating a new security problem.
The agent on my own site
Let me start with something personal.
When I rebuilt saddammalik.com, the site went out in eight languages: English, Hindi, Simplified Chinese, Arabic, Malay, Russian, Spanish and Turkish. My clients and community are spread across Europe, the US, the Middle East and Asia, and many people who find me are more comfortable in their own language.
I couldn't answer every visitor personally at 3 AM in eight languages. So I added Zyra, an assistant that answers questions about my work and replies in the visitor's language, including Hinglish, the Hindi-English mix a lot of my Indian audience actually types in.
I deliberately started Zyra narrow. Today it answers from a curated knowledge base built only from what I have published on my site, and it points people to the official contact page when they're ready. It doesn't trade, take payments or promise anything. A model connection comes next, and it will stay inside the same limits.
That limited scope is exactly why it's useful. The best agents I've seen are the ones with a clear, narrow job, and the safest way to ship one is to start with the boundaries, then add intelligence inside them.
Agents that genuinely help Web3 projects
1. Community and support agents
Telegram and Discord communities never sleep, and neither do the questions: "How do I claim?", "Is this link official?", "Why is my transaction pending?"
A support agent connected to your docs, FAQ and announcements can answer the repetitive questions instantly, in many languages, and hand the hard ones to a human moderator.
The biggest benefit is safety. An agent can tell every user, every time, that the team will never DM first and will never ask for a seed phrase. Scammers love busy communities, and a consistent, always-on reminder helps protect them.
2. Research and summarising agents
Crypto moves fast. Teams drown in governance proposals, competitor announcements, on-chain data and news. A research agent that gathers sources, summarises them and flags what changed is a real time-saver for founders, analysts and content teams.
The rule here: sources first. An agent that summarises without showing its sources will eventually invent something, and in crypto one wrong "fact" can move a community.
3. Trade-assist agents (not trade-for-you agents)
This one is close to my heart, because I've been trading since 2011.
I build agents and bots that assist traders: alerts on Telegram, rule-based signals with risk controls, portfolio summaries, backtesting helpers. What I'm careful about is agents that are given open access to move funds on their own.
I've watched exchanges collapse, protocols get hacked and the largest liquidation day on record in October 2025. Markets can go wrong in minutes. Any agent that touches money needs strict limits: position caps, allow-lists, human approval for anything unusual, and logs of every action. If you can't explain exactly what your agent is allowed to do with funds, it's allowed to do too much.
4. Internal automation with LLMs and RAG
Some of the most valuable AI work isn't flashy. It means turning manual work into automated flows with GPT-class models and your own data, using retrieval-augmented generation (RAG):
- Drafting first versions of reports, KOL briefs and campaign recaps.
- Answering internal questions from a private knowledge base.
- Classifying and routing support tickets.
- Preparing listing-application documents from existing materials (with a human always reviewing).
These don't make headlines, but they save teams hours every week.
5. On-chain aware agents
Agents that can read on-chain data (balances, events, contract state) and explain it in plain language are powerful for dashboards, analytics and user support. "Your transaction is pending because gas is low; here's what to do" is a better experience than a raw block explorer link.
Where I'm careful
Some ideas sound exciting but need extra caution:
- "Fully autonomous" agents with wallets. Possible, but most teams don't need full autonomy. They need speed with guardrails.
- Agents that promise returns. If an agent is marketed as "AI that makes you money," that's a red flag, not a feature. Nobody honest can guarantee returns.
- Agent tokens without an agent. If the token launches before a working product exists, ask why.
- Agents trained on private user data without clear consent and protection.
For ambitious ideas like multi-agent systems and reasoning prototypes, I treat the work as R&D: experiments with clear goals, an evaluation report and a roadmap to production. Honest experiments, not hype.
How I build them
The process is the same one I use for every development project:
- A short call. You share the goal, and I share an honest opinion, including "you don't need an agent for this" if that's the truth.
- A written scope and a fixed quote. No surprise invoices.
- Weekly demos on a staging link. You see the agent answering real questions early, not on launch day.
- Launch with dashboards and logs. You can see what the agent said and did, and why.
- Handover. You own the code, prompts, data pipelines and keys. Nothing is locked to me.
On the technical side, I typically connect modern LLM APIs to your tools and data (docs, databases, on-chain reads, Telegram or X), with retrieval for accuracy and strict permissions for anything sensitive. The stack is chosen for your goal, not for fashion.
Why a trader-developer?
I'm not only a developer. I've been in the market for fifteen years, and I've collaborated with AI and data projects such as Vana, Privasea, Infinity Ground and Kaito, alongside exchanges, wallets and infrastructure. That combination helps in two ways:
- I understand what users actually need from an agent, because I'm one of those users.
- I understand what can go wrong, because I've seen it go wrong, many times, with real money.
A simple test before you build
Ask yourself three questions:
- What exact job will this agent do, and for whom?
- What is it NOT allowed to do?
- How will we know it's working? (Questions answered, time saved, tickets resolved, users helped.)
If you can answer all three clearly, you're ready to build something genuinely useful.
If you're a Web3 team thinking about an AI agent, chatbot or LLM automation for your site, Telegram or X, see what I offer on my AI agents page, read about the Web3 infrastructure & AI projects I've worked with, or book a call. You can also try Zyra on the site and ask it about my work.
Saddam Malik is a crypto trader since 2011, an independent Web3 and AI agent developer, and a crypto KOL. Brand names are trademarks of their owners and describe past collaborations. Nothing here is financial advice.
Originally published at saddammalik.com.
Top comments (0)