Executor

In the plan-and-execute pattern, the executor is the partner to the planner. The planner decides what to do; the executor does it. For each step in the plan, the executor selects the right tool, fills in arguments, runs the call, captures the result, and decides whether to continue, retry, or escalate. In simple single-agent systems the executor and planner are the same LLM running a ReAct loop. In multi-agent systems the executor is often a separate sub-agent with its own narrower toolset and prompt, optimized for action rather than reasoning.

The executor is where most things break in production. Tool calls fail, APIs return unexpected shapes, network requests time out. A good executor handles these failures with explicit policies: retry with exponential backoff for transient errors, escalate to the planner for logic errors, fall back to a default value for non-critical fields. The executor should also report progress back upstream (to the planner or orchestrator) after every step, not just at the end. Without progress reporting, a multi-step task that fails halfway gives you no signal about how far it got.

You separate the executor from the planner when the cost of running tools is high, when steps need different specialized tools, or when you want to enforce a strict approval gate between planning and execution.

• LangGraph plan-and-execute
• Anthropic: Building Effective Agents

Related Terms