Declarative agents

Model

• state: external to the agent, world model
• solution space: all possible states
• agent <-> tool (act) <-> state
  • category theory: focus on the properties of the mappings and their compositions, not the state itself
  • validation is an observe tool, also a human input
• the goal of the agent is to change the state to meet validation criteria within given the constraints
• agent explores the solution space by generating trajectories (sequences of tool calls)
• intelligence is trajectories optimisation

Meta Model

• agents are generated by agents, and their trajectories are observations
• same properties:
  • constraints, e.g., correct syntax
  • validations, e.g., trajectories that lead to goal

Properties

• agent is not an asset, it's an ephemeral entity constructed as needed to guide LLM to solve a problem
• agency is defined by set of tools available to agent that determines all possible trajectories in solution space
  • the more tools are availble to an agent the bigger is the solution space
• act tools change the external state, function calls with effects

Process

• decompose complex tasks into set of agents modelled as state, constraints, validations and tools

Examples

• TDD: given a broken test suite change code until it works
• Hanoi towers and a million steps
• Weather agent