How AI Agents Turn One-Line Prompts Into Fully Functional Apps in Minutes

Overview of ChatDev

ChatDev is a SaaS‑level intelligent software development platform that enables users to generate complete applications by providing only a short natural‑language prompt. The platform advertises a development cycle of approximately three minutes and a cost of about two yuan per app.

One‑Line Demo: "Red Envelope Rain" Game

The article illustrates the process with a simple "red envelope rain" click‑game. The user supplies a single sentence describing the game mechanics, and the AI agents automatically design, code, test, and deliver the functional software.

"Red envelope rain" is a click‑based mini‑game where red rectangular envelopes fall from the top, spaced evenly without overlap; clicking an envelope makes it disappear and shows a monetary value.

During the entire workflow, roles traditionally filled by a product manager, programmer, designer, and tester are all performed by AI agents.

Multi‑Agent Workflow (ChatChain)

ChatDev implements a "ChatChain" architecture, a chain of specialized AI agents that communicate via dialogue to carry out software engineering tasks. The chain follows a waterfall‑style sequence: requirements analysis, design, implementation, integration testing, and documentation.

Users input a project name and a prompt; an optional "one‑click refinement" feature can improve poorly written prompts automatically.

Underlying Large Model: CPM‑Cricket

ChatDev is powered by CPM‑Cricket, the third‑generation 100‑billion‑parameter Chinese Pre‑trained Model (CPM) developed by the same company. CPM‑Cricket shows significant improvements in logical reasoning, code generation, knowledge, and instruction understanding, surpassing Llama 2 in benchmark evaluations.

Benchmark results on standard LLM test sets (HumanEval, C‑Eval, MMLU, MBPP, CMMLU, BBH) demonstrate the model’s strong performance.

Application Example: Luca 3.0 Evaluation

Luca 3.0, a ChatGPT‑like product built on CPM‑Cricket, was tested on a public‑exam dataset (425 questions covering common sense, quantitative reasoning, data analysis, logical inference, and language comprehension). The system answered questions quickly and provided detailed step‑by‑step solutions, achieving accuracy comparable to or exceeding GPT‑4 on key reasoning tasks.

Training Strategies

Curriculum Learning (CL) : The model is first trained on easy tasks to learn basic reasoning, then progressively on harder tasks to align with human reasoning patterns.

Chain‑of‑Thought (CoT) : The reasoning process is decomposed into intermediate steps, improving interpretability and accuracy.

Infrastructure Frameworks

BMTrain – efficient large‑model training framework.

BMInf – high‑performance inference engine.

BMCook – model compression toolkit.

The company reports integration of over 16,000 real‑world APIs and an Int8 quantized version of the model that reduces inference cost by about 50%.

Open‑Source Projects

Beyond ChatDev, the team has released two major open‑source initiatives:

AgentVerse – a general platform for building AI agents with perception, reasoning, collaboration, and execution capabilities.

XAgent – an application framework that gives agents autonomous planning and decision‑making abilities.

These projects, together with ChatDev, form the company’s “three‑horse carriage” focused on AI agents.

Vision: Internet of Agents (IoA)

The long‑term goal is to create an "Internet of Agents" where AI agents interconnect with physical objects and digital twins, enabling a higher‑dimensional "Intelligent Network" that enhances productivity and interaction.

Company Background

Founded in August 2022, the startup’s leadership includes a former Google China employee as CEO and a Tsinghua professor as chief scientist. The team collaborates with OpenBMB and the Tsinghua NLP Lab, contributing open‑source models (CPM‑Ant, CPM‑Bee 10B) and acceleration tools (BMTrain, BMCook, BMInf, OpenPrompt, OpenDelta).

While the vision of a fully connected AI agent network remains aspirational, the released products and open‑source contributions represent concrete steps toward that future.