Working on agents

What I built at TikTok since early 2025, next to what the industry was doing.

2025

In the industry

Early 2025

I paid for Devin, and it was the first time I really used an agent. AI was no longer just models. Of OpenAI's 5 levels of AI, the first 2, chatbots and reasoners, were work for the people who train models. Level 3, agents, was arriving, and as a software engineer I could build it.
So I left TikTok's campaign work and conventional software engineering behind and went all in on agents. I started TikTok's internal general-purpose agent, first on my own with LangChain and the newly announced MCP.
The ReAct paper led me to CodeAct, the research behind OpenHands: an agent acts by writing and running code. I rebuilt the demo on OpenHands. It planned, browsed, ran code, and searched the web, and showed each step as it ran.

Devin is the showpiece for autonomous agents: it finishes the job instead of answering.
Andrej Karpathy names “vibe coding”, and Claude 3.7 Sonnet ships with a preview of Claude Code.
Manus launches and everyone rushes to study it. OpenManus, an open-source take, appears within days.

Apr – Jun

Connected the agent to TikTok's operations data and actions through MCP.
Pushed the OpenHands version further: a new UI, browsing, file upload, and MCP tools the agent could write for itself.

Manus takes off, and agents like it appear: Genspark's Super Agent and the open-source Suna.
Codex arrives on the web. Claude 4 ships and Claude Code becomes generally available.
Dia opens its beta.

Jul

After Manus, open-source general agents like Suna set the pattern. We moved our agent onto Suna's framework as the base for production. I added Kimi K2 3 weeks after its release, and we ran self-hosted open models alongside hosted ones.
Surveyed the agent infrastructure stack: sandboxes (E2B, Daytona, Modal), browsers built for agents (Browserbase, Browser Use), and model routing (OpenRouter).
Started a knowledge base, inspired by NotebookLM. Weighed retrieval approaches from plain RAG and services like Vectorize to Microsoft's GraphRAG, and explored long-term memory with mem0.
Worked on prompt optimization with tools like Dify, and traced agent runs end to end with Langfuse.
Studied how agents are evaluated: SWE-bench and Multi-SWE-bench, GAIA, τ-bench, EvalPlus, Chatbot Arena, and Video-MME for video.

Moonshot open-sources Kimi K2, a model built for agentic tool use.
ChatGPT agent combines research and action. Dia opens a skill gallery for reusable browser workflows.

Aug

Added data tools for creators (profiles and side-by-side comparisons) and slide generation.
Evaluated the ways an agent can get information from the web: search APIs (Exa, Tavily, Brave) and crawlers and readers (Firecrawl, Jina Reader). Built web and image search on Brave, plus YouTube search.
Started moving from one agent that does everything toward capabilities people can pick.

Sep

Shipped skills in our agent, inspired by Dia: reusable workflows that people create, test, version, publish, and share.

Oct

Rendered interactive widgets within days of the Apps SDK. I built the widget host, widget state that survives a reload, and fullscreen widgets.
Added video analysis for a single video or a whole batch.

OpenAI launches apps in ChatGPT and the Apps SDK.
Cursor ships Plan Mode. Days later, Claude Code adds an interactive question tool, so the agent asks before it guesses.
Anthropic introduces Agent Skills and the SKILL.md format.

Nov – Dec

Built plan mode for our agent, inspired by Cursor and Claude Code. The agent proposes a plan, and the user reviews it before anything runs.
Moved the chat to a new thread interface and built a set of TikTok widgets: video lists, comment analysis with CSV export, creator profiles, search, and command output.

Gemini 3 leans into generative UI. Cursor 2.1 lets the agent ask clarifying questions through an interactive UI.
Agent Skills become an open standard.

In the industry

Jan – Feb

Moved our skills to SKILL.md.
Made runs pause for a person: when a tool needs the user to sign in or approve, the run stops and resumes where it left off. Durability became a design constraint for our cloud agents.
My first merge from a coding-agent branch.

Mar – Apr

Started a desktop agent app, choosing pi as its agent core. Within weeks the entry point looked settled: Codex kept getting stronger, and Doubao Work was taking shape inside the company. We stopped the app at the end of April and stopped building a general agent.
After OpenClaw took off, CLIs became how agents reach tools. I contributed preview-deployment commands to TikTok's internal CLI, packaged Lark (TikTok's Slack), operations data, and deployment as CLIs and installable skills, and wrote a skill that teaches agents to design agent-first CLIs.
Built a deploy CLI that coding agents install and use themselves: generated code becomes a shareable URL in one command. This was the start of Sites.

Codex, Claude, and Doubao Work take shape as general agents, and people start their work there.

May – Sep

Sites grew from that CLI into its own product, TikTok's Sites, a month before ChatGPT Sites launched. Any team's coding agent builds a Site already wired into employee sign-in, internal data, Lark, and AI models, from a quick demo page to a production back office. The deploy CLI reached 1.0 and became part of Sites.
Every Site became agent-native. Each can bind its own agent and exposes its user actions over MCP and WebMCP, so every Site can also be a ChatGPT plugin. New Sites get a database automatically, plus file storage and scheduled tasks.
Agents can publish a Site, read its logs, query its data, and roll it back through the same APIs people use.

Anthropic's Thariq Shihipar argues that HTML beats Markdown for agent output, and the idea spreads fast.
OpenAI launches ChatGPT Sites in public beta: ChatGPT builds, publishes, and hosts websites and lightweight apps.
WebMCP lets agents in the browser call a page's actions directly.
OpenAI DevDay's program includes a session on Sites and the next generation of web apps.