The operator and agents

On the desktop and the CLI, Perch is not only a chat surface. It is an operator: it can take actions on your files, run real analysis, and break a large task into coordinated steps. This is the depth behind the web chat, and the reason Perch can produce a finished deliverable rather than only a suggestion.

The operator loop

When you give Perch a task on the desktop or CLI, it works in a loop. It inspects the relevant files, decides what to do next, takes an action, and reads the result before deciding again. The actions available to it include reading and searching across a folder, running commands, and producing structured output.

This is what separates doing the work from describing it. Perch can open the actual files, find the actual numbers, and act on what is there rather than estimating from a summary.

Running analysis in a sandbox

For work that needs real computation, such as auditing a set of payments or reconciling a statement, Perch writes and runs analysis code in a sealed sandbox. The sandbox runs without network access and is isolated from the rest of your machine, so the analysis is contained.

The benefit is accuracy. Rather than guessing whether two payments are duplicates, Perch can run the check and show you the result, along with the records behind it. The analysis is visible in the transcript, so you can see what was run, not just what was concluded.

Focused workers on larger tasks

A large task is rarely one step. For bigger jobs, Perch can plan a sequence and hand parts of the work to focused workers, each with its own narrow context. This keeps each step clear and lets the work proceed in pieces rather than as one overloaded request.

A few principles keep this reliable:

• The driver stays fixed. The main model for a task is chosen at the start and does not change mid task, so the work stays consistent.
• Workers are isolated. Each worker handles a bounded piece with its own context, which keeps one step from polluting another.
• You stay informed. The steps are visible as the work proceeds, rather than hidden behind a single opaque answer.

Two working styles

The operator runs underneath both working styles:

• Saffron is the operator at full intensity: reading files under pressure, running analysis, and finding the thing that matters in a folder.
• Quill points the same engine at writing: producing drafts that are structured and backed by sources.

Both share one workspace and one memory. The difference is the kind of work, not the underlying engine.