Construct capable agents, give them real work, and run them reliably.
This path is about deciding what responsibility an agent should have, what machinery it needs, and how to supervise it once it is doing real work. The outcome is a useful, well-supervised agent, not more agents or more autonomy for its own sake.
The categories move from building and extending an agent, to giving an agent a real responsibility, to running a system that keeps working on its own, to running models on your own hardware.
Build and extend an agent → give it a real job → keep it running safely → run models on your own hardware
Start any category from the problem in front of you; only Claude Code: Builder Mode is meant to be followed in order. The results described below are things you build in your own projects with what you learn, not files included with membership. A resource offers downloads only when it says so.
Pi Agent
Pi Agent is about owning and extending agent machinery rather than only configuring an existing assistant: extensions, custom tools, orchestration, and execution you can inspect.
Questions it explores
- When does a real job outgrow what a default tool can do?
- What does custom control cost in engineering and maintenance?
- How do I keep a custom agent inspectable as it grows?
Something you could work toward: A small custom tool or extension that gives an agent one capability your work needs, with execution you can trace.
Start here if you are comfortable reading and changing code, and a default agent keeps falling short on a specific job.
Good to know: Part of the value is knowing when a simpler tool is the better choice.
Hermes Agent
Hermes Agent explores an ongoing specialist relationship: an agent with a clear responsibility, useful context, tools, memory, supervision, and improvement over time.
Questions it explores
- Which responsibility in my studio would an agent actually help with?
- When does an existing agent system earn its place, compared with a coding agent, an always-on runtime, simple software, or my own judgment?
- How do I inspect the result and decide whether it helped?
Something you could work toward: One recurring responsibility, such as a weekly research summary, handed to an agent and judged on the quality of what it returns.
Start here if you have a recurring job that is not coding and wonder whether an agent could own it.
Good to know: It goes beyond installation: the focus is choosing the responsibility and judging the result.
OpenClaw: Agent Operations
OpenClaw: Agent Operations is about the always-on side of agents: work that waits, watches external state, stays available, and has to recover safely when nobody is chatting with it.
Questions it explores
- Which permissions, schedules, and integrations does an always-on agent need?
- How do I observe it, interrupt it, and recover when something goes wrong?
- Which work is safe to leave running, and which needs a person?
Something you could work toward: An agent that watches for one kind of event, acts within clear permissions, and tells you what it did, with a way to stop it.
Start here if you want an agent to keep working while you are away and need it to stay safe and observable.
Good to know: Resources document real setups, including what did not work, rather than promising that every studio should run one.
Local AI
Local AI follows the changing boundary between what you can own and run on your own hardware and what cloud systems still do better. Hardware, memory, capability, privacy, cost, and evidence all shape that judgment.
Questions it explores
- Which work is practical to run locally today, and which is not yet?
- What do privacy and cost actually gain or lose in my case?
- How do I test a local model fairly before relying on it?
Something you could work toward: A fair side-by-side test of a local model against your current tool on a task from your own work.
Start here if privacy, cost, or independence makes local models worth considering. Check the hardware and model requirements of a specific experiment before buying anything.
Good to know: Local models are tested honestly against cloud tools, not presented as their replacement.
Other paths
Move between paths as your project develops; membership does not come with a required order. The Resources guide explains how the library works.
- Build with AI — Improve how you plan, delegate, code, and ship with AI.
- Build Apple products — Turn Apple platform capabilities into thoughtful, reliable products.
- Run the solo studio — Build the operating system and business judgment behind a one-person product studio.
- Market your products — Find the right audience, communicate the product clearly, and build a repeatable path from launch to learning.