Five things I built with AI agents this year
Hi everyone, I'm Tim. This year almost everything I ship is built by talking to AI agents. Here's what came out of it: a couple of products, and a few tools I built for my own workflow that turned out more useful than I expected.
TinyFile — a Telegram bot with three jobs: making Telegram stickers from your videos and photos, one-off AI generations without a subscription, and compressing videos so they're small enough to send anywhere. @tinyfile_bot
HoldSpeak — push-to-talk dictation for macOS. Hold a key, speak, release, and the text lands in whatever field you're in. Runs locally on your Mac and it's free. I built it because talking to AI is 2–3× faster than typing to it. holdspeak.app
AI Kanban — a project manager that lives in the team's Telegram chat. It reads the conversation, turns it into tasks on a Notion board, moves cards as the work actually ships, and can pick up a task itself and hand it back for review. You talk, the board keeps itself up to date.
Auto-reply bot for a hair salon — answers Instagram DMs on its own: services, prices, stylists, finds a free slot and books it. Anything unclear goes to a human admin, who replies straight from Telegram.
Personal tools — a handful of things I built just for myself to take the friction out of my own workflow. Some of them still out of market 👀
The pattern I keep noticing: generating code is the easy part. The hard part is giving agents a way to see their work and get feedback.
How do you review what agents ship?






Comments1
Hi Tim, greetings from Kyiv too. Most of our code is also written with Claude Code, and what made review manageable for us: a separate agent session acts as the tester and clicks through every visible change in a real browser the way a new user would, and its bug reports go straight back to the session that built it. The builder never grades its own work. Anything that touches money or the production database can only be tested on dedicated test accounts. One distribution note: HoldSpeak is something people actively search for (push to talk dictation on Mac, local dictation without a subscription), so a few pages on holdspeak.app per use case and per comparison could bring steady traffic. Do you let agents merge on their own, or do you always read the diff yourself?
1 comment
Hi Tim, greetings from Kyiv too. Most of our code is also written with Claude Code, and what made review manageable for us: a separate agent session acts as the tester and clicks through every visible change in a real browser the way a new user would, and its bug reports go straight back to the session that built it. The builder never grades its own work. Anything that touches money or the production database can only be tested on dedicated test accounts. One distribution note: HoldSpeak is something people actively search for (push to talk dictation on Mac, local dictation without a subscription), so a few pages on holdspeak.app per use case and per comparison could bring steady traffic. Do you let agents merge on their own, or do you always read the diff yourself?