Design Principles

Action Awareness

Enables agents to assess their own state and performance, allowing for moment-by-moment adjustments. This self-awareness allows agents to make moment-by-moment adjustments, correct errors, and optimize their behaviour.

Completeness

Ensuring that all necessary information is included in the task decomposition. This prevents critical aspects from being overlooked.

Human In The Loop (HITL)

Human-in-the-loop design principles might be incorporated to allow for human oversight and intervention in agent processes.

Feedback Loop

Incorporating a Feedback Loop allows agents to learn from their actions and environmental responses, improving their performance over time.

Memory

Agent stores and retrieves conversations, actions, and observations across various timescales. This allows them to learn from experience, maintain context, and make more informed decisions.

Non-Redundancy

Avoiding unnecessary or duplicate content in sub-tasks. This reduces inefficiency and ensures that agents are not performing the same work.

Reward Model

Consideration of a Reward Model is crucial when designing learning agents, as it shapes the agent’s behaviour by defining what constitutes a desirable outcome.

Solvability

Each sub-task should be independently and completely resolvable by at least one agent. This ensures that the system can effectively address all aspects of the problem.

State Management

Effective State Management is necessary for agents to keep track of their plans, completed tasks, and the results of their actions.

Termination Condition

Designing for Clear Termination Conditions is important to ensure agents (especially in multi-agent conversations) know when a task or interaction is complete.