a2a-agentspeak

A2A Agentspeak

AgentSpeak agents on A2A/ACL protocol. (Work in progress).

Features

• Run AgentSpeak agents on an A2A server.
• Describe agent card and skills in an interface file in a dedicated format.
• Targets that are not declared in the interface are private and ignored from incoming messages.
• Relies on A2A-ACL protocol extension that supports tell, achieve, and ask performatives.
• Synchronous answers for ask messages (to consult a belief).
• Asynchronous answers for achieve messages (to request some actions and optionally answer later).
• (Temporary) Two repositories : one to keep track of running agents (hot repository for hot agents) and one to keep track of agents that can be run locally on demand (cold repository for cold agents).
• (Temporary) Hot agent repositories can be accessed by requested interface (formatted, see samples/llm_req_manager_with_orchestrator_and_hot_repository) or by natural language (interpreted by a LLM, see samples/llm_req_manager_with_orchestrator_and_repository_and_nl_selection).
• (Temporary) Cold agent repositories can be accessed by interface (see tests/ping_agent_on_cold_repo, tests/cold_agent_on_repo_with_holes, and tests/spawn).
• (Temporary) Hot agent repositories can receive failure reports and degrade the reputation of failing agents. That reputation is taken into account when selecting agents (see samples/llm_req_manager_with_orchestrator_and_hot_repository and samples/llm_req_manager_with_orchestrator_and_repository_and_nl_selection).
• Configuration of agents with customized actions at init time (see tests/customizable_robot, tests/action, and tests/cold_agent_on_repo_with_holes).
• Different codecs can be used to encode/decode the content of the messages (AgentSpeak agents can encode the content of the message differently than Python agents or Java agents).

Examples and Documentation

Some examples are given in the samples and tests directories.

To run the simple example from samples/ping, first run run_receiver_agent.py , then run run_sender_agent.py . That example does not require access to an LLM.

Requirements

This module relies on the A2A Protocol (package a2a-sdk), the A2A-ACL package (acl-a2a), and python-agentspeak (package agentspeak).

Optional Requirements

To run the examples that use an LLM, you need:

• the package mistralai or litellm and set your MISTRAL_API_KEY in the environment,
• the package openai and set your OPENAI_API_KEY in the environment.

General architecture