5 lessons from running AI agents in production

Demos are easy. Agents that run every day, for real people, on real channels — that's where the lessons live.

What

Five hard-won lessons from shipping Mona (my Telegram second-brain) and Bengkel AI (an automotive Q&A bot, 88 vehicles, live). Not theory — things that broke in production and what fixed them.

Why it matters

A demo agent and a production agent fail in completely different ways. The demo answers a question. The production agent meets thousands of unexpected inputs, spends money, posts publicly, and has to not embarrass you at 3am. The gap between the two is where most projects die.

Who it's for

Developers building real AI agents — not chatbots for a screenshot. If your agent will touch real users, real channels, or real actions, these apply.

When & where

Every one of these surfaced once the agent left the happy path: a weird user message, a tool called in an order I didn't expect, a token budget blown mid-conversation, a channel I hadn't tested.

How

The five lessons:

1. Safety at the agent layer, not just the tool. Gate any side-effecting action before it runs; require approval for anything that spends money or posts publicly.
2. Patterns must be specific, not single keywords. Real intent lives in phrases and context. A sloppy filter is worse than none — it lulls you into trusting it.
3. Multi-channel is the product. The brain is channel-agnostic; the channel is an adapter. Build it that way and one agent ships to Telegram, WhatsApp, web, and voice without a rewrite.
4. Memory + autonomy make it feel alive. Cross-day memory and scheduled actions (a morning briefing, a daily check) turn a chatbot into a teammate.
5. Tokens are a budget. Compression, prompt caching, and on-demand tool schemas turned a slow, expensive agent into a fast, cheap one.

The takeaway

The prompt is maybe 10% of a real agent. The other 90% is plumbing — channels, memory, safety, scheduling. That's exactly why I'm packaging it into a kit.