v0.1.3 Release

v0.1.3

Features

• verl training backend — Support verl as a training backend, with GRPO config, dataset/loop manager, LLM server, trainer, and a GSM8K math-agent example (#70); plus setup and launch documentation (#73). Fire all tasks in the batch concurrently instead of capping by semaphore (#75)
• slime training backend — Add SlimeRunner as the Python entry point for training (#56), with rollout/reward/gateway/trace integration and SGLang token-ID capture patches; Sample.metadata is passed as the agent payload verbatim
• Structured JSON logging — Attach JSON logging with sessionId to the root logger (#63)

Bug Fixes

• Client — Use a unified rate limiter for all ACR API calls (#51)
• Examples — Fix env setup flag (#53); standardize app naming and fix reward-format incompatibility (#39, #41)

Examples

• AppWorld example — New strands_appworld_agent example for AppWorld API interaction tasks (#38); adapt the official baseline system prompt and few-shot examples for good performance (#48)
• tau-bench example — New strands_taubench_agent example, using OpenAIModel for the assistant

Documentation

• Introduce a Starlight documentation site (#65)
• Add a CloudWatch session-logs skill (#71)
• Migrate docs from vLLMModel to standard OpenAIModel (#40)
• Update rLLM integration status and add training-example links (#45)

Full Changelog: v0.1.2...v0.1.3