The trouble with working with agents
Permissions, mistakes, costs and why you have to keep an eye on them: what has actually happened to me.
Yes, there are free AI agents, especially open-source ones like OpenClaw, which you can install on your machine without paying for a licence. But the…
Yes, there are free AI agents, especially open-source ones like OpenClaw, which you can install on your machine without paying for a licence. But the agent is just the shell: the model that does its thinking costs money if it's a paid AI, unless you hook it up to a model you run locally yourself. Completely free, then, only when you supply the computer.
The catch with "free" is consumption. An agent uses far more than a chat, because it goes round and round: every step it thinks through and every check is one more query to the model. If you connect it to a paid AI through the API (the pay-as-you-go access that programs use), the bill can climb without you noticing. With a local model you don't pay per use, but you need a machine with enough memory, and the models that fit at home tend to fall short on long tasks. Some platforms include agents in their free version, but with limits.
I use both paid AI and local models, with Ollama, LM Studio or llama.cpp, and cost is one of the problems I go through in "The problems of working with agents", along with the permissions you give without realising and silent errors. If you want to try without spending, my advice is to set a spending limit on the API from day one.
Permissions, mistakes, costs and why you have to keep an eye on them: what has actually happened to me.
One email a week with what was verified, what was refuted and what is still on the bench. No news roundups.