LLM examples

The LLM integration chapters each introduce one technique. The projects in the LLM examples group show larger applications and the places where LLM reasoning can enter an ASTRA system.

Projects in this guide

Download	Contents	Walkthrough
reasoning-agents.zip	One agent for each reasoning strategy	The reasoning architecture
a2a-greeter.zip	An ASTRA agent published as an A2A service	Publishing an A2A service
a2a-caller.zip	An ASTRA agent that calls the greeter service	Calling A2A services

Projects in the LLM examples group

Project	What to explore
Pipelines	Compares the reasoning strategies on the same problems. The reasoning agents were developed here before moving into `astra-langchain4j`.
A2A	A `MultiStepChainOfThought` solver published as an A2A service, a `ReAct` coordinator that calls it, a scripted model for running without an API key, interoperability tests with non-ASTRA A2A agents, and a pinned copy of the A2A Inspector.
Travel Planner	Specialist agents exchange requests and results in a sequence controlled by a round-robin orchestrator.
Tic-Tac-Toe	LLM-backed test-goal plans choose moves, while the game integration manages the board and turn-taking. The project compares basic, reflective and defensive strategies.
Towerworld	A composite prompt describes the task and block state. The model proposes actions, and ASTRA executes them through EIS and waits for their effects.

The Pipelines and A2A projects use the ASTRA 2.5.3 libraries. The Travel Planner, Tic-Tac-Toe and Towerworld projects were written for the earlier API described in the astra-langchain4j paper, including provider modules named OpenAI and Gemini. Follow each project’s own version and setup instructions; updating its POM alone does not migrate its source code.

The travel planner is the simplest of the earlier projects. Its README explains how to provide the key and run the agents. Tic-Tac-Toe additionally requires its game server. Towerworld uses an EIS environment and other project-specific dependencies; inspect its POM and source before running it.

The ASTRA layer provides the control structure around model calls: it can schedule interactions, check proposed actions, retain knowledge, and recover from failed steps. Use the Failure Handling, Test Goals, Interaction Protocols and EIS guides alongside these projects.