Observation and Action Spaces

Observation and Action Spaces form the interface between the model and the environment – analogous to ISA in computer architecture.

  • Information that does not fall into the observation space does not exist for the model.
  • An operation missing from the action space can only be described in words.

Therefore, with a fixed model, the main lever for improving the agent is to expand context and tools. Many “model problems” are actually interface problems.

Example: Manus combined the Deep Research, coding, and Computer Use spaces; OpenClaw expanded them to the user’s digital life.

Related: Agent Formula, Context as the Agent’s Eyes