is it free to build an ai agent?

what is genuinely free

• Frameworks. Langgraph, crewai, autogen, the openai and anthropic agent sdks, smolagents. All free, all open source, all installable in a minute.
• Local models. Ollama plus an open-weight model runs on a laptop with enough memory and costs nothing per token. Quality is below the hosted frontier models, and for narrow tasks that is often fine.
• Free tiers. Most providers give you a small credit to start. Enough for a prototype, not enough for anything with users.

So the honest answer to the literal question is yes, you can build something that works this afternoon for nothing.

what is not free

That bottom row is the entire story. Everything above it is rounding error next to engineering time, and it is the row nobody puts in a comparison.

where the time actually goes

Not the agent loop. The loop is about eighty lines. The time goes to everything around it:

1. Failure handling. Tool call fails, model returns malformed arguments, api times out, agent loops forever. Every one of these needs a policy, and you discover most of them in production.
2. Permissions. What the agent may touch, and how you stop it touching the rest. This gets serious the moment it has write access to anything.
3. Memory. Storing what it learned, retrieving the right part, and pruning so context stays affordable.
4. Cost control. Step budgets, spend caps, and model routing so a single bad run does not bill you $200.
5. Observability. When output is nondeterministic, reconstructing what happened without a trace is close to impossible.

A working demo is a weekend. Something you would leave running against production data is a quarter, and that is with an engineer who has done it before.

the build-versus-buy line

Build when the agent is your product, when the workflow is unusual enough that no platform fits, or when you need to run it in an environment nobody else can reach.

Buy when the agent is a means to an end. The list above is undifferentiated work that every team rebuilds identically, and rebuilding it is the most expensive way to learn what a permission model should look like.

how this works in aldena

Aldena is the buy side of that line, and the free tier exists so the comparison is cheap to make: 2 rooms and 5 agents per room at no cost, so you can run real work before deciding. Paid plans start at $99 a month for 20 rooms and 10 agents each, with no per-agent fee.

Model usage stays separate and visible, billed as credits at published rates, so the cost line you would have built yourself is already itemised. The five items above, failure handling, permissions, memory, spend control, and watching it happen, are the product rather than your backlog.

For the how rather than the cost, see how can i create my own ai agent.