Dual-Mode Hosting
Two ways to build an employee
Every agent in the factory lives inside the same perimeter, draws from a metered budget, keeps its notebook in the safe, and passes an activation smoke check before going live. You choose how much you want to own: a declarative configuration on the shared runtime, or a dedicated repository in its own container.
Default Track · Shared Runtime
Configuration-Only Agent
For most assistants, analysts, and workflow automatons, no git repository or container build is needed. You define the agent as pure configuration running on the factory’s shared agent runtime:
- Persona (Soul): Write the instructions, tone, and boundaries that define who the employee is.
- Memory Archetype: Choose purpose-built (stateless task worker), personal assistant (per-person notebook), or team member (shared project workspace).
- Composable Skills: Attach reusable Python skill packages checked offline by the Registrar.
- Fleet-Wide Upgrades: Approve a new shared runtime digest once and the factory smoke-tests and upgrades every configuration agent automatically.
- Portable Bundle: Export the entire agent configuration as a single bundle and import it into any factory.
Full Control · Custom Container
Repository Cartridge
When an employee needs custom execution logic or specialized runtime dependencies beyond the shared loop, build a self-contained repository cartridge:
- Custom Execution Loops: Implement bespoke multi-step state machines, custom context assembly, or specialized reasoning pipelines.
- Any Language or Binary: Write the cartridge in Python, TypeScript, Go, or Rust, or bundle compiled native libraries and command-line tools.
- Custom Domain Schemas: Maintain specialized local SQLite tables and indexing pipelines inside the agent’s notebook directory.
- House Staff & Platform Tools: Power privileged factory house staff and custom engineering workflows admitted from a pinned commit.
Decision Guide
When to choose each model
Both models are governed by the same ten cast members and pass the same activation smoke gate before taking traffic.
| Dimension | Configuration-Only Agent | Repository Cartridge |
|---|---|---|
| What you author | Persona, memory archetype, caller roles, and adopted skill packages. | Git repository with manifest, persona, source code, unit tests, and Dockerfile. |
| Creation & updates | Instant via API or conversationally with the Agent Builder; updates version in place. | Push a 40-character commit SHA; admitted through offline secret scan, test run, and image build. |
| Tooling & dependencies | Modular Python standard-library and vendored pure-Python skill packages shared across the fleet. | Any language, runtime, or compiled binary tool (never cloud or model provider SDKs). |
| Runtime maintenance | One admin-approved runtime rollout upgrades and smoke-tests the fleet with automatic rollback. | Repository owner maintains the base image, dependencies, and turn loop. |
| Portability | Self-contained configuration bundle export and import across factories. | Git repository and container image; can also be converted in place to a configuration agent. |
Lifecycle
From idea to sleeping employee
Track A
Building a Configuration-Only Agent
- Define persona & memory: Author the soul and select the memory archetype (purpose-built, personal assistant, or team member).
- Adopt skills: Pick approved skills from the catalog, or package and register new Python skills verified offline by the Registrar.
- Create & set policy: Create the agent via the API or Agent Builder, then have an administrator grant its models, routes, and budget with the Accountant and Gatekeeper.
- Smoke test & sleep: Deploy the agent. The Property Manager runs an activation smoke check—verifying a model turn and every adopted skill—before putting the agent to sleep ready for work.
Track B
Building a Repository Cartridge
- Scaffold the cartridge: Write the soul, manifest, turn loop, Dockerfile, and unit tests against the factory model API and local notebook directory.
- Admit a pinned commit: Register the commit SHA. The Registrar and Property Manager scan for hardcoded secrets, run the test suite, and build the container image.
- Bind policy & credentials: Bind required secret names with the Keymaster and grant models, routes, and budget limits.
- Smoke test & sleep: Deploy the admitted image through the activation smoke gate. If a cartridge later simplifies to standard skills, convert it in place to a configuration agent without losing its identity, ledger history, or memory.
Developer Guides
Read the full specifications
Explore the step-by-step guides for creating configuration agents, packaging reusable skills, and authoring custom cartridges.