Smart Agents

[Cormanz]

This aims to resolve #25 by proposing a new Smart Agent solution described below, where an agent is divided into three sub-agents.

Brainstormer

Memory:

• The task and role.
• 15 recent observations.
• 3 guidelines from the Reflector.

Initialized With:

• The last request it gave, the Actor's response.

Action Cycle:

• Extracts 2 or 3 observations from the Actor's response.
• Brainstorms a new idea.
• Gives that idea to the Actor.

Reflector

Called at the start, and after every two Brainstormer-Executor cycles.

Memory:

• The task and role.
• 25 recent observations, weighted for importance as well.
• 8 recent reflections, weighted for importance asa well.

Initialized With:

• The last three guidelines.
• The Brainstormer's ideas during the two cycles.
• The Executor's responses during the two cycles.

Start of Program:

• Think of long-term ideas for how to complete the task, insert these as observations.
• Assign The Brainstormer with three guidelines.

Action Cycle:

• Evaluate whether or not The Brainstormer and Executor are making progress.
• If necessary, reflect over how to complete the task from here, long-term, generalizing all observations and reflections.
• Assign The Brainstormer with three new guidelines.

Executor

Memory:

• 5 recent observations.

Initialized With:

• The task given by the Brainstormer.

Action Cycle:

• Turn that task into a runnable command.
• Turn the results of that command into human-readable output.