DECAF provides a stable way of designing, rapid development and execution of agents to achieve solutions in particularly complex software systems. It includes modules that perform all the necessary services of an intelligent agent: communication, planning, scheduling, execution monitoring, coordination, and finally, learning and self-diagnosis. The main ones are outlined below.

• Agent initialization
  
  This module runs when an agent is started. It reads the plan file and each task reduction specified in the plan file is added onto a plan library (the Task Templates Hash Table), together with a tree structure that specifies the actions that will achieve that objective.


• Dispatcher
  
  Control is passed to this module once the Agent Initialization has done its job. The dispatcher waits for an incoming message (written in KQML or FIPA) - see Communication. The messages are placed on the Incoming Message Queue. It can perform one of two tasks: either finding the corresponding action specified by the message in the Pending Action Queue (if the message is part of an existing conversation), or in the other case, it creates a new objective and places it on the Objectives Queue for the Planner. Needless to say, not all active objectives can be achieved.


• Planner
  
  The Planner looks at the Objectives Queue and matches new goals to an existing task stored in the Plan Library. A copy of the plan (once it is instantiated) is placed in the Task Queue, with a unique identifier and any other provisions passed to the agent by means of the incoming message. If another message with the same goal does arrive, then another instantiation of the same plan will be placed with a new unique identifier. So the Task Queue contains instantiated plans that need to be completed in response to an unsatisfied message.


• Scheduler
  
  This waits until the Task Queue isn't empty, so that it can determine which actions can be executed at a given time, which should be executed, and the order in which this should happen. This can only be done if all the provisions for a particular module are available at the time. Some of these come from the incoming message, and some as a result of other completed actions. So the Task Queue needs to be continuously checked each time a provision becomes available to decide which actions can be executed at what time.


• Executor
  
  This module starts to work when the Agenda Queue isn't empty: as soon as an action is placed on this queue, the executor carries out the task. Either the action completes normally, in which case the results are placed on the Action Result Queue, and so the framework waits for the results, distributes them to actions further down that may be waiting in the Task Queue. Once this is done, the module will check the queue for further work, and will start each task individually, keeping up with deadlines. The alternate case would be if the action fails and therefore doesn't return, so the framework will indicate failure to the requester.

Building the DECAF framework does have one main objective, and that is to enable very rapid development of agent actions and task structures. For this to be achieved, the programmer shouldn't have anything to do with agent interaction details. The developer doesn't need to write any communication code, or worry about many invocations of the same agent; no API is needed to write and program an agent under the DECAF architecture.

DECAF does involve two middle agents (see Middleware) as well. One of them is Matchmaker, which helps agents to find other agents in the community that may provide a service that is useful for their tasks. An agent will "advertise" its capabilities with the Matchmaker and if these change or are no longer available, the agent will delete them. The Matchmaker then stores this information in a local database, so that requesters can formulate queries to it and "ask" for a set of matching advertisements, or even "subscribe" to it and hence be informed when new services of interest are either added or removed.

The other agent is the Broker, which advertises with the Matchmaker, an overview of its capabilities built from the providers that have registered with one Broker. So for example, a Broker may have all the capabilities needed to build a house (plumber, electrician, roofer,…), so it can provide a larger service than any single provider can, and can also manage a large group of agents more effectively.

Currently, DECAF is the basis for AI (Artificial Intelligence) projects, as well as those involving organizational development and bioinformatics information gathering. It has also been used as a programming tool for agent development.