Continually and autonomously evolving agent system

Continually learns new capabilities from tasks, autonomously extending the Agent system.

Apple Silicon · macOS 13+Windows 10 / 11 · x64

Weekly experiment report Field research Autonomous updates2

Compile the three result sets in experiments/ into a weekly report, and export the charts as PNG.

Processed2m 14s

Load skill · Read ×2 · Python code +1

Load skill weekly-report Read run_1.csv run_2.csv run_3.csvPython codeGranted

The three result sets have been summarized into a weekly report, the charts exported as PNG and written to the project folder; this workflow was saved as a skill.

Message CLAgent…

Default

Background

When a neural network learns new tasks one after another, it overwrites capabilities it has already acquired, that is, catastrophic forgetting(French, 1999). How to let capabilities keep accumulating as tasks continue is the core problem of continual learning(Parisi et al., 2019). CLAgent brings it into the agent system: it learns continually while carrying out tasks, and system capabilities expand accordingly.

French, R. M. (1999). Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4), 128–135.

Parisi, G. I., et al. (2019). Continual lifelong learning with neural networks: A review. Neural Networks, 113, 54–71.

Autonomous evolution by layer

All three layers are evolving; every change takes effect after a decision.

MemoryWrites, rewrites and forgets facts and preferences with each session; later tasks retrieve and reuse them
AppsGeneralizes skills and rewrites its own apps: proposal, dry run on a copy, becoming a new generation
SystemLoads and decides on changes from the two layers above; apart from system calls, all other functions evolve autonomously

Continual learning loop

System capabilities

Files, terminal, browser and model are provided by the kernel; new capabilities are added as apps.

Files and documents

Reads and writes files in the workspace, and generates Word, PowerPoint and spreadsheet files.

Terminal and code

Runs shell commands and Python, with the output visible line by line.

Browser

Opens, reads and acts on web pages in the built-in browser.

Process and permissions

Every step is recorded on the timeline; sensitive operations are approved first.

Model

Official service, or connect an OpenAI-compatible endpoint.

Apps

One capability, one app; it can be installed, replaced and evolved.

Usage

Two sources for models; the difference is whose gateway forwards the model requests.

Official service

Model requests are forwarded through the official gateway and metered by token

Available models are maintained by the official service
Usage is recorded in real time and can be checked in Account

Your own endpoint

Model requests go directly to the user's own service endpoint

Works with OpenAI-compatible endpoints
The model list and parameters are configured by the user