PeopleBot in plain English

Imagine you want an assistant to check a project’s instruction sheet. You want to know which version it read, what it found, and whether another assistant can follow up later. PeopleBot describes that work using five main ideas. The names are technical, but the responsibilities are familiar.

​

Blueprint: the reusable job description

A Blueprint describes how a kind of agent behaves. Think of it as a job description you can reuse. It does not contain every project’s files or give an agent access to someone’s accounts. For example, a review Blueprint can describe the required evidence and how findings should be reported. Your project and task are supplied separately.

​

Instance: the particular worker

Two reviewers can use the same Blueprint and still be different Instances. Each has its own identity and selected memory. Reusing instructions does not copy someone else’s work history or credentials.

​

State: the exact saved version

Git saves snapshots called commits. Each commit has a long identifying code. A State reference names the repository, the full commit, and optionally a path inside that snapshot. This lets “the file I reviewed” mean a specific saved file, even if someone edits it afterward. In contrast, “the latest file” can change between the request and the review. Try the snapshot lesson to see the difference.

​

Message: the handoff

The intended message is a small request or reply that points to the necessary State and results. It helps another actor find the work without receiving an entire private conversation. The public 0.1.0a3 package does not implement message delivery. Message references in a record describe part of the design; they do not create a working inbox. See the coordination design.

​

Execution: one attempt

An Execution records one attempt at a task: its objective, exact inputs, and outcome. A retry is another attempt, not a reason to rewrite the history of the first one. An answer, a saved result, and a remotely shared result are separate facts. Recording each fact accurately helps someone decide what to do after a failure.

​

Two supporting ideas

An environment is the owning setup around the agents: its computer/runtime, identities, repositories, and credentials. It decides what work is allowed. An adapter connects PeopleBot to a particular runtime. Think of it as the connector that knows how to invoke that runtime and interpret its response. It is not another agent or a sixth core concept.

​

What memory means here

Instance memory is selected information saved for future work: a decision, a task checkpoint, or a useful procedure. It is not the model’s hidden thought process and does not automatically include every earlier conversation. A future Execution can read the exact saved items it needs. The caller currently chooses and saves those items explicitly; the public release does not learn or summarize them automatically. See memory and recovery.

​

Try it yourself

Start with the beginner learning path. You will use ordinary files before working with agent configuration. For the original technical definitions, see the public framework overview.