A Planner hands each task to an Executor, which calls one of several tools through a typed schema. A Critic checks the tool's output against the plan; if it fails, the task loops back for another attempt, up to a retry limit, otherwise it is dropped.
P(accept) = reliability * (1 - strictness)
retry while attempts < maxRetries, else drop
- Task spawn rate — how often the Planner emits new work.
- Tool reliability — base chance a tool call returns a usable result.
- Critic strictness — how much extra scrutiny the Critic applies before accepting.
- Max retries — attempts allowed before a task is dropped as failed.
Accepted tasks flow into the green Memory node and are cached; rejected tasks flash amber and loop back to the Planner, or flash red and vanish once retries run out.