The Multiplexer is responsible for maintaining potentially multiple connections.


Connections wrap an external SDK or API and manage messaging. As such, they allow the agent to connect to an external service with an exposed Python SDK/API.

The module connections/ contains the abstract class which defines a Connection. A Connection acts as a bridge to the SDK or API to be wrapped, and is, where necessary, responsible for translating between the framework specific Envelope with its contained Message and the external service.

The framework provides one default connection:

  • stub: implements an I/O reader and writer to send messages to the agent from a local file.

InBox and OutBox

The InBox and OutBox are, respectively, queues for incoming and outgoing Envelopes. They are needed to separate the thread which runs the Multiplexer from the thread which runs the main agent loop.


An Envelope is the core object with which agents communicate. It travels from OutBox to another agent or gets translated in the Connection to an external service or protocol. Envelope objects sent from other agents arrive in the InBox via a Connection. An Envelope is a vehicle for messages with five attribute parameters:

  • to: defines the destination address.

  • sender: defines the sender address.

  • protocol_id: defines the id of the protocol.

  • message: is a bytes field which holds the message in serialized form.

  • Optional[context]: an optional field to specify routing information in a URI.


Protocols define how messages are represented and encoded for transport. They also define the rules to which messages have to adhere in a message sequence.

For instance, a protocol may contain messages of type START and FINISH. From there, the rules may prescribe that a message of type FINISH must be preceded by a message of type START.

The Message class in the protocols/ module provides an abstract class with all the functionality a derived Protocol message class requires for a custom protocol, such as basic message generating and management functions and serialisation details.

The framework provides one default protocol:

  • default: this protocol provides a bare bones implementation for an AEA protocol which includes a DefaultMessage class and a DefaultSerialization class with functions for managing serialisation. Use this protocol as a starting point for building custom protocols.

Additional protocols can be added as packages, including:

  • oef_search: this protocol provides the AEA protocol implementation for communication with the OEF search node including an OefSearchMessage class for hooking up to OEF search node services and search agents. Utility classes are available in the module which provides OEF search node specific requirements, such as classes, needed to perform querying on the OEF search node, such as Description, Query, and Constraint, to name a few.
  • fipa: this protocol provides classes and functions necessary for communication between AEAs via a variant of the FIPA Agent Communication Language. For example, the FipaMessage class provides negotiation terms such as cfp, propose, decline, accept and match_accept.


Skills are a result of the framework's extensibility. They are self-contained capabilities that AEAs can dynamically take on board, in order to expand their effectiveness in different situations. A skill can be given permission to read the internal state of the the AEA, and suggest action(s) to the AEA according to its specific logic. As such, more than one skill could exist per protocol, competing with each other in suggesting to the AEA the best course of actions to take.

For instance, an AEA who is trading goods, could subscribe to more than one skill, where each skill corresponds to a different trading strategy. The skills could then read the internal state of the AEA, and independently suggest profitable transactions.

A skill encapsulates implementations of the abstract base classes Handler, Behaviour, and Task:

  • Handler: each skill has none, one or more Handler objects, each responsible for the registered messaging protocol. Handlers implement AEAs' reactive behaviour. If the AEA understands the protocol referenced in a received Envelope, the Handler reacts appropriately to the corresponding message. Each Handler is responsible for only one protocol. A Handler is also capable of dealing with internal messages.
  • Behaviour: none, one or more Behaviours encapsulate actions that cause interactions with other agents initiated by the AEA. Behaviours implement AEAs' pro-activeness.
  • Task: none, one or more Tasks encapsulate background work internal to the AEA.

Skills further allow for Models. Classes that inherit from the Model can be accessed via the SkillContext.


Main loop

The _run_main_loop() function in the Agent class performs a series of activities while the Agent state is not stopped.

  • act(): this function calls the act() function of all active registered Behaviours.
  • react(): this function grabs all Envelopes waiting in the InBox queue and calls the handle() function for the Handlers currently registered against the protocol of the Envelope.
  • update(): this function dispatches the internal messages from the decision maker to the handler in the relevant skill.

Decision maker

The DecisionMaker component manages global agent state updates proposed by the skills and processes the resulting ledger transactions.

It is responsible for the AEA's crypto-economic security and goal management, and it contains the preference and ownership representation of the AEA.

TransactionMessage and StateUpdateMessage

Skills communicate with the decision maker via InternalMessages. There exist two types of these: TransactionMessage and StateUpdateMessage.

The StateUpdateMessage is used to initialize the decision maker with preferences and ownership states. It can also be used to update the ownership states in the decision maker if the settlement of transaction takes place off chain.

The TransactionMessage is used by a skill to propose a transaction to the decision maker. The performative TransactionMessage.Performative.PROPOSE_FOR_SETTLEMENT is used by a skill to propose a transaction which the decision maker is supposed to settle on chain. The performative TransactionMessage.Performative.PROPOSE_FOR_SIGNING is used by the skill to propose a transaction which the decision maker is supposed to sign and which will be settled later.

The decision maker processes messages and can accept or reject them.


For examples how to use these concepts have a look at the `tac_` skills. These functionalities are experimental and subject to change.


Filter routes messages to the correct Handler via Resource. It also holds a reference to the currently active Behaviour and Task instances.

By default for every skill, each Handler, Behaviour and Task is registered in the Filter. However, note that skills can de-active and re-activate themselves.

The Filter also routes internal messages from the DecisionMaker to the respective Handler in the skills.


The Resource component is made up of Registries for each type of resource (e.g. Protocol, Handler, Behaviour, Task).

Message Envelopes travel through the Filter which in turn fetches the correct Handler from the Registry.

Specific Registry classes are in the registries/ module.

  • ProtocolRegistry.
  • HandlerRegistry.
  • BehaviourRegistry.
  • TaskRegistry.