The framework has two distinctive parts.

  • A core that was originally developed by the Fetch.ai team, that has now been forked by Valory-xyz.
  • Extensions (also known as packages) developed by any developer.

Currently, the framework supports four types of packages which can be added to the core as modules:

  • Skills encapsulate logic that deliver economic value to the AEA. Skills are the main focus of the framework's extensibility.
  • Protocols define the structure of agent-to-agent and component-to-component interactions (messages and dialogues) for agents.
  • Connections provide interfaces for the agent to connect with the outside world. They wrap SDKs or APIs and provide interfaces to networks, ledgers and other services.
  • Contracts wrap smart contracts for Ethereum and third-party decentralized ledgers.

The following figure illustrates the framework's architecture:

Simplified illustration of an AEA

The execution is broken down in more detail below:

Execution of an AEA

The agent operation breaks down into three parts:

  • Setup: calls the setup() method of all registered resources
  • Operation:
    • Agent loop (Thread 1 - Asynchronous agent loop):
      • react(): this function grabs all Envelopes waiting in the InBox queue and calls the handle() method on the Handler(s) responsible for them.
      • act(): this function calls the act() method of all registered Behaviours.
      • update(): this function enqueues scheduled tasks for execution with the TaskManager and executes the decision maker.
    • Task loop (Thread 2- Synchronous): executes available tasks
    • Decision maker loop (Thread 3- Synchronous): processes internal messages
    • Multiplexer (Thread 4 - Asynchronous event loop): processes incoming and outgoing messages across several connections asynchronously.
  • Teardown: calls the teardown() method of all registered resources

To prevent a developer from blocking the main loop with custom skill code, an execution time limit is applied to every Behaviour.act and Handler.handle call.

By default, the execution limit is set to 0 seconds, which disables the feature. You can set the limit to a strictly positive value (e.g. 0.1 seconds) to test your AEA for production readiness. If the act or handle time exceed this limit, the call will be terminated.

An appropriate message is added to the logs in the case of some code execution being terminated.