Course contentsShow
← Machine Learning and Deep Learning
Lesson 2069 of 3,53844. AI Agents: FoundationsFree lesson

Single-Turn vs. Multi-Turn Agents

Compare agents that solve tasks in one step versus those that iterate multiple times to refine solutions.

Single-Turn vs. Multi-Turn Agents

What you'll learn: How agents differ based on whether they complete tasks in one action versus iterating through multiple steps to refine their approach.

The Core Distinction

In the agent world, single-turn agents execute their entire task in one shot—they perceive the environment, decide on an action, execute it, and finish. Think of them as sprinters: they analyze the situation once, take their best action, and cross the finish line.

Multi-turn agents, by contrast, operate through multiple cycles of the perception-action loop. They take an action, observe the result, reason about what happened, then decide on the next action. They're like chess players, adjusting strategy after each move based on how the board changes.

When Each Approach Works

Single-turn agents excel when:

  • The task is straightforward and deterministic
  • All necessary information is available upfront
  • One well-chosen action suffices (e.g., classifying a query and routing it to the right retrieval system)

Multi-turn agents shine when:

  • Tasks require exploration or verification
  • Initial attempts may fail and need correction
  • Complex reasoning demands breaking problems into steps
  • The environment provides feedback that informs subsequent actions

The Trade-off

Single-turn agents are faster and cheaper—one LLM call, one action, done. Multi-turn agents are more robust and capable—they can recover from mistakes, handle uncertainty, and tackle problems requiring iterative refinement. However, they consume more tokens and time with each additional turn.

Many real-world agentic systems blend both: they might use multi-turn reasoning for complex queries but fall back to single-turn responses for simple questions.

Key Takeaway: Single-turn agents optimize for speed with one-shot execution, while multi-turn agents trade efficiency for adaptability through iterative perception-action cycles.