Good Life Games

Agent workflow prototype

Shepherd

A local-first control room for structured agent work.

Shepherd connects a task backlog, approval points, verification gates, run records, logs, and artifacts.

The workflow

Define the task. Run the work. Require evidence.

Each handoff is explicit, and every completed run keeps the results needed to review it.

01Choose a task

Read the backlog and its dependencies.

02Approve the plan

Accept, reject, or change the approach.

03Run one iteration

The agent works on one eligible task.

04Execute checks

Typecheck, tests, build, and browser tests run as gates.

05Save the record

Results, logs, and artifacts stay with the run.

06Review the evidence

Inspect what passed, failed, or needs another run.

Working interface

Tasks and verification stay visible.

Shepherd task backlog with status, priority, and dependency information
Structured task backlogThirty-four tasks with status, priority, and dependencies.
Shepherd run detail showing passed typecheck, test, and build gates
Verified runA fresh local run with typecheck, test, and build results plus artifact paths.

The larger idea

The run record connects agent decisions, automated checks, and human review.

That same pattern now informs the broader Automation Playbook.

Explore the playbook →