Planning

Higher-order reasoning about reasoning processes, including strategy selection and monitoring

Automated planning approach that decomposes complex tasks into hierarchically structured networks of simpler tasks using domain knowledge

Systematic task decomposition, progress tracking, and adaptive workflow management for complex multi-step processes

Systematic breakdown of complex objectives into achievable, measurable sub-goals with clear success criteria

Planning approach that simultaneously satisfies multiple constraints while optimizing objectives

Multi-scenario planning that prepares for different possible futures and uncertain conditions

Generates the full plan or tool-call graph up front, decoupled from tool observations, and only then executes it. ReWOO writes a variable-parameterized chain of tool calls before anything runs, cutting redundant model invocations; LLMCompiler builds a dependency DAG and dispatches independent tool calls in parallel. This differs from ReAct, which interleaves one thought and one action at a time and re-feeds every observation into the prompt.

The agent keeps an explicit, editable todo or plan artifact (for example a write_todos tool call or a todo.md file) and continually rewrites and recites it as work proceeds. Re-emitting the evolving plan at the end of the context window pushes the global objective back into the model recent attention on every step, which resists goal drift and lost-in-the-middle forgetting across long runs.

The agent maintains, or uses the LLM itself as, a world model that predicts the next state and consequence of each candidate action, then runs look-ahead rollouts ("imagine before acting") to score and select an action before executing anything in the real environment. This extends model-based reinforcement learning to language agents and matters most where actions are irreversible and backtracking on a live website or GUI is impossible. Distinct from scenario-planning, which reasons across several strategic futures with no action-consequence simulator, and from reflective-mcts, which searches over reasoning traces rather than simulated environment states.