Skills are the core focus of the framework's extensibility as they implement business logic to deliver economic value for the AEA. They are self-contained capabilities that AEAs can dynamically take on board, in order to expand their effectiveness in different situations.
A skill encapsulates implementations of the three abstract base classes
Model, and is closely related with the abstract base class
Handler: each skill has zero, one or more
Handlerobjects, each responsible for the registered messaging protocol. Handlers implement AEAs' reactive behaviour. If the AEA understands the protocol referenced in a received
Handlerreacts appropriately to the corresponding message. Each
Handleris responsible for only one protocol. A
Handleris also capable of dealing with internal messages (see next section).
Behaviour: zero, one or more
Behavioursencapsulate actions which further the AEAs goal and are initiated by internals of the AEA, rather than external events. Behaviours implement AEAs' pro-activeness. The framework provides a number of abstract base classes implementing different types of behaviours (e.g. cyclic/one-shot/finite-state-machine/etc.).
Model: zero, one or more
Modelsthat inherit from the
Modelsencapsulate custom objects which are made accessible to any part of a skill via the
Task: zero, one or more
Tasksencapsulate background work internal to the AEA.
Taskdiffers from the other three in that it is not a part of skills, but
Tasks are declared in or from skills if a packaging approach for AEA creation is used.
A skill can read (parts of) the state of the the AEA (as summarised in the
AgentContext), and suggest actions 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. In technical terms this means skills are horizontally arranged.
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 preference and ownership state of the AEA, and independently suggest profitable transactions.
The framework places no limits on the complexity of skills. They can implement simple (e.g.
if-this-then-that) or complex (e.g. a deep learning model or reinforcement learning agent).
The framework provides one default skill, called
error. Additional skills can be added as packages.
Independence of skills¶
horizontally layered, that is they run independently of each other. They also cannot access each other's state.
Two skills can communicate with each other in two ways. The skill context provides access via
self.context.shared_state to a key-value store which allows skills to share state. A skill can also define as a callback another skill in a message to the decision maker.
The skill has a
SkillContext object which is shared by all
Model objects. The skill context also has a link to the
AgentContext provides read access to AEA specific information like the public key and address of the AEA, its preferences and ownership state. It also provides access to the
This means it is possible to, at any point, grab the
context and have access to the code in other parts of the skill and the AEA.
For example, in the
ErrorHandler(Handler) class, the code often grabs a reference to its context and by doing so can access initialised and running framework objects such as an
OutBox for putting messages into.
Moreover, you can read/write to the agent context namespace by accessing the attribute
Importantly, however, a skill does not have access to the context of another skill or protected AEA components like the
What to code¶
Each of the skill classes has three methods that must be implemented. All of them include a
teardown() method which the developer must implement.
Then there is a specific method that the framework requires for each class.
There can be none, one or more
Handler class per skill.
Handler classes can receive
Message objects of one protocol type only. However,
Handler classes can send
Envelope objects of any type of protocol they require.
handle(self, message: Message): is where the skill receives a
Messageof the specified protocol and decides what to do with it.
A handler can be registered in one way:
- By declaring it in the skill configuration file
It is possible to register new handlers dynamically by enqueuing new
Handler instances in the queue
context.new_handlers, e.g. in a skill
component we can write:
Behaviour class contains the business logic specific to initial actions initiated by the AEA rather than reactions to other events.
There can be one or more
Behaviour classes per skill. The developer must create a subclass from the abstract class
Behaviour to create a new
act(self): is how the framework calls the
A behaviour can be registered in two ways:
- By declaring it in the skill configuration file
- In any part of the code of the skill, by enqueuing new
Behaviourinstances in the queue
context.new_behaviours. In that case,
setupis not called by the framework, as the behaviour will be added after the AEA setup is complete.
The framework supports different types of behaviours:
OneShotBehaviour: this behaviour is executed only once.
act()method is called every
tick_interval. E.g. if the
TickerBehavioursubclass is instantiated
There is another category of behaviours, called
SequenceBehaviour: a sequence of
Behaviourclasses, executed one after the other.
FSMBehaviour: a state machine of
Statebehaviours. A state is in charge of scheduling the next state.
If your behaviour fits one of the above, we suggest subclassing your
behaviour class with that behaviour class. Otherwise, you
can always subclass the general-purpose
Follows an example of a custom behaviour:
from aea.skills.behaviours import OneShotBehaviour class HelloWorldBehaviour(OneShotBehaviour): def setup(self): """This method is called once, when the behaviour gets loaded.""" def act(self): """This methods is called in every iteration of the agent main loop.""" print("Hello, World!") def teardown(self): """This method is called once, when the behaviour is teared down."""
If we want to register this behaviour dynamically, in any part of the skill code (i.e. wherever the skill context is available), we can write:
Or, equivalently to the previous two code blocks:
The callable passed to the
act parameter is equivalent to the implementation
act method described above.
The framework is then in charge of registering the behaviour and scheduling it for execution.
Task is where the developer codes any internal tasks the AEA requires.
There can be one or more
Task classes per skill. The developer subclasses abstract class
Task to create a new
execute(self): is how the framework calls a
Task class implements the functor pattern.
An instance of the
Task class can be invoked as if it
were an ordinary function. Once completed, it will store the
result in the property
result. Raises error if the task has not been executed yet,
or an error occurred during computation.
We suggest using the
task_manager, accessible through the skill context,
to manage long-running tasks. The task manager uses
schedule tasks, so be aware that the changes on the task object will
not be updated.
Here's an example:
from aea.skills.tasks import Task def nth_prime_number(n: int) -> int: """A naive algorithm to find the n_th prime number.""" assert n > 0 primes =  num = 3 while len(primes) < n: for p in primes: if num % p == 0: break else: primes.append(num) num += 2 return primes[-1] class LongTask(Task): def setup(self): """Set the task up before execution.""" def execute(self, n: int): return nth_prime_number(n) def teardown(self): """Clean the task up after execution."""
from aea.skills.behaviours import TickerBehaviour from packages.my_author_name.skills.my_skill.tasks import LongTask class MyBehaviour(TickerBehaviour): def setup(self): """Setup behaviour.""" my_task = LongTask() task_id = self.context.task_manager.enqueue_task(my_task, args=(10000, )) self.async_result = self.context.task_manager.get_task_result(task_id) # type: multiprocessing.pool.AsyncResult def act(self): """Act implementation.""" if self.async_result.ready() is False: print("The task is not finished yet.") else: completed_task = self.async_result.get() # type: LongTask print("The result is:", completed_task.result) # Stop the skill self.context.is_active = False def teardown(self): """Teardown behaviour."""
The developer might want to add other classes on the context level which are shared equally across the
Task classes. To this end, the developer can subclass an abstract
Model. These models are made available on the context level upon initialization of the AEA.
Say, the developer has a class called
Then, an instance of this class is available on the context level like so:
Each skill has a
skill.yaml configuration file which lists all
Task objects pertaining to the skill.
It also details the protocol types used in the skill and points to shared modules, i.e. modules of type
Model, which allow custom classes within the skill to be accessible in the skill context.
All AEAs have a default
error skill that contains error handling code for a number of scenarios:
- Received envelopes with unsupported protocols
- Received envelopes with unsupported skills (i.e. protocols for which no handler is registered)
- Envelopes with decoding errors
- Invalid messages with respect to the registered protocol
The error skill relies on the
fetchai/default:1.0.0 protocol which provides error codes for the above.
Custom Error handler¶
The framework implements a default
You can implement your own and mount it. The easiest way to do this is to run the following command to scaffold a custom
Now you will see a file called
error_handler.py in the AEA project root.
You can then implement your own custom logic to process messages.