Everyone wants to build an agent right now, and almost nobody agrees on what the word means, which makes pricing it a mess. A demo where a model calls one API and summarizes the result costs almost nothing. An agent that reliably takes actions across your real systems, handles the cases where a tool fails, and doesn't do something expensive when it misreads a request is a genuine engineering project.
The gap between those two is where all the cost lives. Here's how to think about what an AI agent actually costs to build, what moves the number, and the honest tiers. Duskel builds agents from $2k for a tightly scoped one, with the more ambitious multi-step systems running as retainers from $3k a month.
The two things that decide the price
The first is how many tools the agent touches and how good those tools are. Every action an agent can take is a tool: search the database, send the email, create the ticket, issue the refund. Each one needs a clean interface the model can call without guessing, error handling for when the underlying system says no, and permissions so it can't do something it shouldn't. Two tools is an afternoon. Twelve tools across five systems that were never designed to talk to each other is the project.
The second is the blast radius of a mistake. An agent that drafts a reply for a human to approve can be wrong occasionally and it costs nothing. An agent that sends the reply, moves money, or changes a customer's account has to be right, and "has to be right" is where the money goes: validation, confirmation steps, audit logs, rollback, and the testing to prove it behaves under the weird inputs real users provide. Read-only agents are cheap. Agents that act are not.
The cost tiers, honestly
Simple: a single-purpose assistant that reads and reasons but doesn't take irreversible actions. It answers questions against your data, drafts content, triages incoming requests, or routes things to the right place. One or two tools, low stakes, a human in the loop for anything that matters. This is the from-$2k end, and honestly it's where a lot of teams should start.
Mid: a multi-step agent that chains several tools to get something done and takes real actions with guardrails. It might process a support case end to end, enrich and update records, or run a workflow that used to need a person. Now you need orchestration, retries, state that survives a failed step, and the confirmation logic that keeps it from compounding one wrong decision into ten. This is a monthly engagement because the edge cases surface over weeks of real traffic, not in a two-week sprint.
Complex: multiple agents, or one agent operating in a high-stakes domain with little human oversight. Long-running tasks, coordination between specialized agents, tight integration with systems where errors are costly. This is ongoing retainer work, and the cost reflects that most of the effort is spent on the failure modes, not the happy path everyone saw in the demo.
The ongoing cost people forget
An agent is not a build-once asset. Models change, your systems change, and the ways users try to break it are endless and creative. Agents drift. Something that worked in March starts doing something dumb in June because a downstream API changed its response format or a new model version reasons differently. Budget for monitoring and maintenance the same way you'd budget for keeping a hire effective, not the way you'd budget for buying a desk.
There's also the running cost of the model calls themselves. A multi-step agent can make many model calls per task, and those tokens add up faster than a single chatbot reply. It's usually manageable, but it's a real per-use number that scales with volume, and it's worth measuring early rather than discovering it on a bill.
Where the money gets wasted
The classic mistake is building an autonomous agent when a workflow would do. If the steps are predictable, you don't need a model deciding what to do next on every task; you need reliable automation with the model doing the one genuinely fuzzy part. Agents are for when the path is unknown, not for dressing up a process you could have scripted for a tenth of the price.
The other trap is skipping the human-in-the-loop stage and going straight to full autonomy. Start with the agent proposing and a person approving. It's cheaper, it builds the trust and the data you need, and it stops you from paying to clean up an expensive mistake before you'd even proven the thing works.