Tools & agents
Agentic Workflow
Also known as: agentic AI
A way of using AI where the model plans and carries out a multi-step task with some autonomy, deciding its own next actions, rather than answering in one shot.
An agentic workflow is when AI works through a task in steps, deciding what to do next as it goes, instead of giving one immediate answer. Ask it to "research this topic and write a report", and it might search, read, take notes, draft and revise, choosing each move itself. It is "AI that gets things done" rather than "AI that just replies", and it is one of the biggest shifts in how AI is being used.
An agentic workflow is a task carried out by an AI agent with a degree of autonomy: the model plans, takes actions using tools, observes results, and decides the next step, looping until the goal is met.
Agentic vs fixed
- Fixed workflow: the steps are predefined by a developer (orchestration decides the path).
- Agentic workflow: the model decides the path at runtime, so it can adapt to what it finds.
More autonomy means more flexibility but less predictability. In practice, systems mix the two, and it is often wiser to use the least autonomy a task needs.
Why it is a big deal
It moves AI from answering questions to completing work, powering coding assistants, research assistants and automation, but reliability, cost and oversight become the central challenges.
"Agentic" describes systems where an LLM exercises runtime control over its own trajectory, selecting tools and next actions in a perceive-reason-act loop, as opposed to developer-fixed orchestration. The distinction is a spectrum of autonomy; influential guidance frames a workflows vs agents choice and recommends the minimum autonomy sufficient for the task, since model-driven control trades reliability for flexibility (Anthropic 2024).
Core ingredients: planning/decomposition, tool use, memory (short-term context, longer-term retrieval), and iteration with feedback. The dominant engineering problems are reliability under compounding error, cost/latency of long loops, safe handling of side-effecting actions, and evaluation of open-ended trajectories, mitigated by bounded loops, structured output, least-privilege tools, verification/critique steps, and human-in-the-loop checkpoints. Agentic workflows scale to multi-agent topologies when tasks are decomposable or benefit from independent verification.