What the term covers
Agentic AI is an umbrella, and its edges are drawn differently by different vendors. What every definition shares is one shift: the model's output stops being the end of the interaction. It becomes an action — a search, a database query, a code change — or the input to the next model call, and the system keeps going until a goal is met or a limit is reached.
Anthropic's guide to building these systems, published in December 2024, uses agentic systems for the whole family and splits it in two. In a workflow, the developer writes the path: call the model, then a tool, then the model again, in an order fixed in code. In an agent, the model chooses the path itself, step by step. IBM frames the same territory from the top down — agentic AI as a goal-directed system, often of several cooperating agents under an orchestration layer, working with limited supervision. Both descriptions point at the same practical change: software that acts, not just software that answers.
Agentic AI versus an AI agent
The two terms are often used interchangeably, but the distinction is useful. An AI agent is one kind of system: a model that decides its own next step in a loop. Agentic AI is the wider category that includes agents and everything around them — fixed multi-step pipelines, systems of several agents handing work to each other, and the orchestration that ties them together.
A three-step pipeline that drafts, checks and formats a document is agentic but is not an agent, because code decides every step. A coding assistant that reads a failing test, decides which file to open and keeps trying until the test passes is both. The difference shows up on the bill: a fixed pipeline costs the same number of calls every time, while an agent's cost depends on how many steps it decides to take.
The common patterns, and what each does to cost
Anthropic's guide names a short list of building blocks that most agentic systems combine. Each has a recognisable cost shape:
- Prompt chaining — split a task into fixed sequential calls, with checks in between. Cost is predictable: a known number of calls, each carrying the previous step's output.
- Routing — classify the input first, then send it to a specialised prompt or model. This is the one pattern that can lower cost, by sending easy cases to cheaper models; see model routing.
- Parallelisation — run independent parts at once, or run the same task several times and compare. Faster, but every parallel call is billed.
- Orchestrator and workers — one model plans and delegates subtasks to others, then merges their results. Adds a planning call and a merging call on top of the work itself.
- Evaluator and optimiser — one model drafts, another critiques, repeat until it passes. Cost is open-ended unless the number of rounds is capped.
- Autonomous agent — a model using tools in a loop on its own judgement. The least predictable cost of all, because the model sets the number of steps.
Why agentic systems cost more than chat
A chat reply is one model call. An agentic task is many, and they are not independent: each call carries the instructions, the tool definitions and everything that has happened so far. A model keeps no memory between calls, so the history is resent and billed again at every step. The result is a cost that is mostly input and grows faster than the number of steps.
Anthropic is plain about the trade in its own guide: these systems spend more time and money in exchange for better results on hard tasks, and that exchange should be weighed before building one. The worked example on this page prices one task three ways — a single call, a fixed chain and an agent loop — on the same model.
Worked exampleOne task three ways: a single call, a fixed chain, an agent loop
1,000 requests a month on Claude Sonnet 5.5, priced at its current list rates, checked on 2026-10-09.
Each “request” here is one whole task, 1,000 tasks a month. The token totals per task are illustrative assumptions: one call; three calls, each carrying the previous step's output; twenty agent steps averaging a 20,000-token context, with the repeated history cached. The agent's cost is almost all input.
Change the workload or the model in the LLM API cost calculator.
Keeping agentic cost predictable
- Use the simplest shape that works. A single call with good retrieval often beats a pipeline; a fixed pipeline beats an agent when the steps are known in advance. Anthropic's own advice is to add complexity only when it measurably helps.
- Cap every loop. A maximum number of steps or review rounds turns an open-ended cost into a bounded one.
- Cache what repeats. Instructions and tool definitions sit at the start of every call; prompt caching bills them at the cached rate after the first time.
- Send routine steps to small models. Planning may need a frontier model; extracting a field or formatting a result rarely does.
- Batch what does not need an answer now. Independent subtasks can run through a batch API at a discount.
- Measure per task. Log steps and tokens per completed task; a per-token price says little about an agentic system on its own.
Where agentic systems go wrong
The defining strength — acting without a human in every step — is also the source of the failures. A wrong assumption at step two is inherited by every step after it, so errors compound rather than average out. IBM's list of risks adds the subtler ones: systems that game the metric they were given, optimisation that runs further than intended, failures that cascade through a chain of cooperating agents, and decisions that are hard to trace after the fact.
The standard defences are structural rather than clever: require human approval before consequential actions such as payments, deletions or messages sent on someone's behalf; run tools in a sandbox; keep a full trace of every call; and test the system on realistic tasks before trusting it with real ones.