0cedac72d718db6c6d018e53fe4e3be9324ff390
- Agent contract: OPENAI_BASE_URL, OPENAI_API_KEY, AGL_TASK_INPUT, AGL_EVENT_URL (was 3, now 4 with auth key) - Algorithm key: always used by algorithm for auth. Optionally stored in K8s Secret — algorithm may run outside cluster and get key from its own config. - Added Authorization headers to all algorithm Python examples in get_started.md
Agl-lite: A Minimal Workable Version of Agent Lightning
This project agl-lite which is a minimal workable version of the popular agentic rl project Agent Lightning (https://github.com/microsoft/agent-lightning).
here are some key changes:
- remove the dependency of
litellmand build self-owned request gateway instead; - remove the dependency of
OpenTelemetryand the whole stack built upon it, such as the tracers of agents; instead, use gateway to collect request-response data during transfer and record them into data store; - following (2), the organization of data is not based on span in opentelemetry, instead, the basic trajectory format is sequence of requests (with response);
- use
kubenetesas the default agent runner (minikubefor single machine), and move the retry control from data store to k8s controller, and simplify the rollout states
There are some local environment setup and configuration needed for this project refactoring in .local/.
Description
用于快速构建 AI Agent 的训练工具。|GitHub 镜像 ⭐ 17.7k · 🍴 1.6k
https://github.com/microsoft/agent-lightning
Readme
MIT
56 MiB
Languages
Python
98.1%
Shell
1.9%