What Is the “Lobster” AI Agent Trend and How Does It Work?

The term “lobster” (小龙虾) has become popular in AI circles to describe a pattern where an AI model runs continuously on a server and is controlled through instant‑messaging platforms like Telegram, Slack, or WeChat bots.

Traditional AI usage follows a short interaction: open a webpage, ask a question, read the answer, and close. In contrast, the lobster approach treats the AI as a long‑lived program that stays active, awaiting commands such as: /fix payment bug When the AI receives a command, it typically:

Looks at the code

Analyzes the problem

Modifies files

Submits a pull request

This workflow feels like remotely directing an AI programmer.

The nickname comes from the community’s joke that the agent constantly crawls code repositories and consumes tokens, much like a lobster scuttling around a server and eating everything in sight.

Technical Building Blocks

Agent framework – enables the AI to call external tools (e.g., LangChain, AutoGen).

Skills (tools) – give the AI abilities such as executing shell commands, reading/writing files, calling APIs, and editing code.

IM bot – receives messages (Telegram Bot, Slack Bot) and forwards them to the agent.

Memory – stores context via markdown documents, vector databases, or a Git repository.

Scheduled tasks – use cron to trigger the AI periodically.

Combined, these components form an Agent + Tools + Scheduler + Chat entry point system.

Why It Suddenly Gained Traction?

Two practical reasons drive its popularity:

AI providers benefit from higher token consumption; a continuously running agent uses far more tokens than occasional queries, turning the model into a service rather than a tool.

The experience feels addictive: seeing AI‑generated progress reports in a team channel gives the impression of a new “digital employee,” which many find appealing.

However, the hype also attracts criticism. Demonstrations often show AI doing part of the work while humans finish it, or AI merely summarizing daily activities—what some call “cyber chanting.”

Realistic Use Cases

Automatic CI fixing – when CI fails, the AI reads logs, proposes a fix, and opens a PR.

Code inspection – scheduled scans of repositories to spot potential issues.

Information aggregation – automatically fetches GitHub updates, papers, or tech news and generates summaries.

These scenarios share traits: they are repetitive, rule‑based, and tolerate occasional errors.

今日进展： 修复2个bug 新增3个测试 发现1个潜在问题

In summary, the “lobster” style simply transforms AI from a chat‑based assistant into a continuously running automation program. It may not be the ultimate solution, but it clearly signals a shift in how AI is being integrated into development workflows.