Each step of an agent plan runs the same loop: the planner picks the next tool, the executor calls it through a typed schema (inputs, outputs, preconditions), the environment returns an observation, and memory stores the result before the next step begins. A whole plan of length n only succeeds if every grounded call along the way does.
P(plan succeeds) = r^n
r = per-call tool reliability, n = planning depth
- Planning depth — how many tool calls the plan chains before it is considered complete.
- Available tools — the toolset the planner can choose from at each step.
- Tool reliability — probability a single grounded call returns a valid observation (schema-checked; the rest fail and are logged, not silently accepted).
- Action speed — how fast the agent travels through the plan-act-observe loop.
The core orb is the agent; beams reach out to the tool it calls, and the ring above stores the fixed-capacity memory of recent observations — full slots evict the oldest, just like a real context window.