Agent
An AI agent uses a model to choose steps, call tools, observe results, and continue toward a goal.
Use an agent when the task needs adaptive planning across uncertain steps; use a workflow when the path is known.
Use when
- Research and synthesis tasks
- Multi-step tool use
- Open-ended support triage
- Operations where the next step depends on observations
Avoid when
- Deterministic approval flows
- High-risk transactions without review
- Simple API orchestration
- Tasks with a fixed known sequence
The agent decision
An agent is useful when the model must choose what to do next. It may search, call tools, inspect results, revise a plan, and continue. This flexibility is powerful, but it creates more failure modes than a fixed workflow.
The strongest early agent use cases are bounded enough to evaluate but flexible enough that hard-coding every step would be brittle.
When to use a workflow instead
If the steps are known, a workflow is usually better. A workflow can still use an LLM at specific points, but the application controls the sequence.
Common mistakes
- Calling every tool-using LLM flow an agent.
- Skipping evaluation because behavior is “emergent”.
- Giving the agent high-impact tools before adding guardrails.
Next decision
Compare agents with workflows first. Then decide what tools, memory, evaluation, and human review the system needs.