Managed API
The managed entry point to Arcie. Product direction; in development, not a launched API reference.
Arcie—Holeacquisition LLC’s model-agnostic agent infrastructure.
The open-source framework and the managed API belong to one agent system. The framework gives developers control of their agent code and execution. The API is being built for developers who want Holeacquisition LLC to operate the agent engine.
In development. This page describes the managed API direction, not a production API contract. No managed Agent API endpoints, SDK methods, pricing, availability dates or service guarantees are announced here. For current framework usage, start with the quickstart.
The intended developer experience
Build the product around your agents. Supply instructions, tools, knowledge and tasks through an API rather than maintaining the agent orchestration service yourself. Installing the Arcie framework should not be a prerequisite.
The managed service is intended to handle the agent loop, session state, tool coordination and connected execution environments. The runtime must distinguish tools hosted in its environment from tools invoked through external services.
One system, two entry points
| Entry point | Developer responsibility | Execution responsibility |
|---|---|---|
| Open-source framework | Agent files, custom tools, application logic and runtime operation | The process running Arcie, locally or in a deployment |
| Managed API (in development) | Agent configuration, connected tools, application logic and permissions | Holeacquisition LLC-operated agent engine and configured execution infrastructure |
Shared agent semantics and a path between these modes are the architectural direction. Automatic migration or identical behavior across deployments is not a current guarantee.
Model-agnostic, not capability-blind
The goal is provider choice without rebuilding the surrounding product. The runtime must still account for differences in tool calling, context limits, reasoning behavior and other provider capabilities. Supporting a model does not mean it can perform every task as well as another model.
Different models for specialized subagents and safe model changes during a long-running task are engineering goals for the managed service, not promised features of the current API.
What needs to be proven
- Long-running tasks continue without an application holding a connection open.
- Runs can be inspected, steered, paused, canceled and resumed.
- Recovery preserves checkpoints and does not blindly repeat completed external actions.
- Tools execute within explicit permissions, isolation and spending limits.
- Usage and execution events support debugging, accounting and end-user billing.
- Model adapters preserve supported capabilities and fail clearly when a capability is unavailable.
Task completion, recovery behavior and cost per successful task must be measured before the managed service is described as production-ready.
How it connects to Holeacquisition LLC
Gateway supplies model access. Sessions supplies turn and event interfaces. Memory supplies persistent context. Compute supplies execution environments. Arcie connects those layers into agent behavior; it does not replace them.
The framework’s existing use of Holeacquisition LLC Sessions is a foundation, not evidence that this managed Agent API is already available.