Seven building blocks. One loop. Everything a language model needs to not just talk, but to act.
The model thinks. A controller behind it, the harness, keeps the loop running, calls tools, and feeds the results back to the model. Round after round, until the task is done.
The loop runs until the task is done, sometimes two rounds, sometimes two hundred.
Every agent is built from these parts. Tap a module. Behind each one are the real tools and models as of October 2026.
The fundamental choice for any project. Toggle to see both sides.
Runnable locally or at a host of your choice, with full control over your data. The price: your own hardware and some setup.
The strongest performance with no effort, but your data flows through the vendor.
You can run an open model yourself: on your own machine, or on a rented one. Renting means you pay for what you use and need no hardware. Running it yourself means you need enough memory — how much depends on the size of the model. The number before the B counts its parameters in billions: 7B is small, 70B is large.
With the wizard below you can assemble the setup your use case needs — and set it up with an AI agent of your choice.
The key terms, briefly explained.