Every frontier model provider enforces a different tool schema, safety prompt, and execution loop. When these API specifications change, custom orchestration wrappers break. For parallel RL loops or automated evaluation suites, these infrastructure discrepancies destroy trajectory determinism.
A clean benchmark harness absorbs this variance instead of forcing the developer to patch it. The goal is a hardened runtime that standardizes provider-specific plumbing and secures the underlying OS interaction, ensuring evaluations measure actual agent reasoning rather than environment variance.
computer-1 is a model-agnostic computer-use agent built directly inside Harbor, the official harness for Terminal-Bench 2.0. It runs 20+ benchmark adapters out of the box using isolated Docker containers locally, and scales out to cloud environments via Daytona and Modal without requiring core agent logic rewrites.
One Interface, Four Providers
Instead of maintaining separate orchestration logic for every vendor, computer-1 standardizes tool calling, safety acknowledgments, and token/cost accounting across four model backends through a single implementation:
To keep the installation footprint clean, the runtime enforces fail-fast validation at construction: an explicit model flag is required, native flavors reject non-computer-use models before making an API call, and vendor SDKs are loaded lazily via the harbor[computer-1] extra.
computer-1 finding Cambridge's highest-energy covered property in the BEUDO dataset, a multi-step browser task requiring map navigation and data extraction across records. 116 steps on Claude Opus 4.8; verifier passes. View full run →
Getting Started
Instead of rebuilding custom tracking frameworks for every new model vendor, computer-1 lets you swap the model string, keep the harness, and maintain an identical trajectory format. This shifts the engineering task from managing vendor-specific environment variance to analyzing agent behavior at scale.
The implementation lives as a self-contained module within harbor/src/harbor/agents/computer_1. You can add it to your environment via pip install 'harbor[computer-1]'.