Guides¶
The concepts say what the parts are. These say how to build with them.
Each guide answers one question and assumes the concepts behind it. Read them in the order below to build an application from nothing, or go to the one that names your question.
How to¶
| Guide | Answers |
|---|---|
| Write an agent | What an agent does when it is woken, what it declares, and what acknowledging commits it to |
| Write an admission rule | How the application refuses a write before the board sequences it |
| End a run | What ends a run, who closes it, and how the outcome is read |
| Serve a blackboard over HTTP | Which path carries each operation, and what your service supplies |
| Notify agents over HTTP | How a notification reaches an agent that is its own service |
| Test an application | How to drive time, agents and a whole run without waiting |
The first three are the application's own code. The next two are what a deployment adds. The last one applies throughout.
Language models¶
An agent decides its own work, and a language model is one way to decide. Nothing in the library depends on either page.
| Guide | Answers |
|---|---|
| Give an agent a language model | How the board's operations become tools a model can call, against any provider |
| Use the board inside LangChain | How those tools join a loop LangChain already runs |
Read the first before the second. The first is this library's own tool surface, against any provider; the second is one framework's way of driving it.