Planning & Task Decomposition

ReAct is powerful but fundamentally linear: think, act, observe, think again, one step at a time. That works well for a research question. It works poorly for "set up the new marketing campaign," where a PM expects the agent to produce fifteen tasks, assign dependencies between them, and execute in a sensible order, not stumble through them one ad hoc thought at a time. This is where plan-and-execute takes over from pure ReAct.

A planning agent splits the work into two distinct phases. First, generate a complete plan: a full breakdown of the goal into subtasks, with the dependencies between them made explicit. Second, execute that plan, running steps in an order that respects those dependencies, and running any steps that have no dependency on each other concurrently rather than one at a time.

Represent the plan as a directed acyclic graph, a DAG: each task is a node, and an edge from task A to task B means B cannot start until A finishes. A campaign might decompose into: draft the brief (no dependencies), design creative assets (needs the brief), write ad copy (also needs the brief, but not the assets), set up tracking pixels (no dependencies at all), and launch (needs everything else done first).

Toggle the failure scenario below and watch the plan actually recompute, not just display a different hardcoded picture.

If a step fails and it genuinely cannot be recovered by simply trying it again, the right response is not to keep retrying it forever, and it is not to abandon the whole plan either. It is to reconsider the remaining plan itself: is there an alternative path that reaches the goal without this step, or with a substitute for it? The code above shows exactly this: when the tracking pixel API is down for good, the Planner does not block the whole campaign launch on it. It drops that dependency and proceeds with a fallback, manual tracking, instead.