Dashboard|

Introduction

Arcie—Holeacquisition LLC’s model-agnostic agent infrastructure. An open-source framework and a managed API in development.

Arcie—Holeacquisition LLC’s model-agnostic agent infrastructure.

An open-source framework and a managed API for building agents that reason, use tools, remember and act. One agent system, with two entry points: operate the framework yourself or use a Holeacquisition LLC-managed service.

Open-source framework

The framework lets you define agents as files. Every part of an agent lives in a predictable place. Arcie discovers the files, assembles a manifest and runs the agent loop, including custom tools and subagents.

npx arcie@latest init my-agent

The idea

An agent is a folder. What it is, what it can do, what it knows, and how it behaves are all expressed as files inside an agent/ directory:

my-agent/
└── agent/
    ├── agent.ts         # model + configuration
    ├── instructions.md  # the system prompt
    ├── tools/           # what it can do
    ├── skills/          # what it knows
    ├── subagents/       # who it delegates to
    ├── channels/        # where it lives (HTTP, Slack…)
    ├── schedules/       # when it acts on its own
    ├── sessions/        # memory + session configuration
    └── policies/        # security, budgets, guardrails

Arcie discovers these slots, validates them, and assembles an agent manifest — the single object the runtime executes. Add a file, get a capability. Delete a file, remove it. Your agent is always exactly what's on disk.

Why filesystem-first

  • No boilerplate. A defineTool in tools/get_weather.ts is a tool. There's nothing to register.
  • Reviewable. Every capability is a file in version control. A diff shows precisely how an agent changed.
  • Composable. Subagents are just nested agent folders. Knowledge is just markdown. Scaling up is adding directories, not rewriting code.
  • Model-agnostic. Connect to supported model providers rather than designing your agent around one vendor. Model and tool capabilities still vary by provider.
  • Connected to Holeacquisition LLC. Use Gateway and Sessions for model routing, conversation state, usage accounting and project controls.

Managed API

We’re building a managed API as another entry point to Arcie, not a competing framework. Developers will supply instructions, tools and tasks while Holeacquisition LLC operates the agent engine and connected execution infrastructure.

The intended boundary is operational responsibility: API users should not need to install the framework or adopt its folder structure. Shared execution semantics and portability between self-operated and managed agents are design goals, not capabilities these docs claim are complete today.

Read the managed API direction.

How it fits with Holeacquisition LLC

Arcie is the agent layer; Holeacquisition LLC supplies the connected infrastructure.

  • Gateway supplies model access and request controls.
  • Sessions supplies conversation state and the turn/event interface used by the framework runner.
  • Memory is the persistent-context layer.
  • Compute supplies environments for deployed agent code and tools.

In the current framework, custom tool code executes in the process running Arcie. Calling the Sessions API does not by itself move that code to Holeacquisition LLC or turn the run into a fully managed, recoverable background task.

Next steps