Dependencies

AgentCompass checks optional Python dependencies only when the active evaluation path needs them. The host is the machine running the current agentcompass process; the generic dependency manager inspects and modifies only the Python interpreter used by that process.

Packages inside a task Environment are prepared by an image, snapshot, Recipe, or harness. System prerequisites such as the Docker daemon, GPUs, and provider credentials are also outside Python dependency management.

Dependency or prerequisite	How it is prepared	Generic auto-install
Python packages in the AgentCompass interpreter	Install an agentcompass[<extra>] optional dependency group	Only for paths marked as supported below
Packages in a harness runner or task Environment	Image, snapshot, Recipe, or component-specific installation	Not applicable
Docker, GPU, system-tool, and provider prerequisites	Operating-system or provider configuration	Not applicable

See Install AgentCompass for the base installation. This page covers only component-specific optional dependencies.

Optional Extras

The table lists optional Python dependency groups installed as agentcompass[<extra>]; <extra> is the name in the first column.

Extra	Purpose	Generic auto-install
swebench	Generate test scripts and parse scores for swebench_verified and swebench_multilingual	Supported
scicode	Run the host-side evaluator for SciCode	Supported
gdpval	Download reference files and provide XLSX report utilities for gdpval_ac	Supported
wildclawbench	Decrypt WildClawBench ground truth on the host	Supported
mini-swe-agent	Run mini_swe_agent with launch_mode=local	Supported only by local mode with install_strategy=install_if_missing
frontier-engineering	Provide openevolve for the OpenEvolve harness	Not supported; prepare it through an image or manual installation
taubench	Run TauBench through host_process	Not supported; prepare it through an image or manual installation

swebench_pro uses its own evaluation scripts and does not require the official swebench Python package on the host.

The scicode extra does not include the additional matplotlib dependency needed by test problem 80; see SciCode data and dependencies for the complete setup.

Manual Installation

To install one or more extras manually from a source checkout:

uv pip install -e ".[swebench,mini-swe-agent]"

For an installed release:

python -m pip install "agentcompass[swebench,mini-swe-agent]"

Automatic Installation

Automatic installation is disabled by default. When it is not enabled, a generic check that finds a missing dependency raises OptionalDependencyError and prints uv and pip commands for the current interpreter. When auto-install is enabled, AgentCompass attempts to install the extra when that check runs and then checks the declared imports again. It prefers uv and falls back to the current interpreter’s python -m pip when uv is unavailable.

agentcompass run <benchmark> <harness> <model> --auto-install-dependencies

Interface	Setting
agentcompass run	Enable with --auto-install-dependencies; explicitly disable with --no-auto-install-dependencies.
agentcompass launch	Use the same option as a shared runtime setting for the entire orchestration.
Configuration or orchestration file	Set runtime.auto_install_dependencies: true.
Python SDK	Pass auto_install_dependencies=True to run_evaluation() or launch(); see the Python SDK.
Environment variable (single request only)	Set AGENTCOMPASS_AUTO_INSTALL_DEPENDENCIES=true; 1, yes, and on are also true, with corresponding false values. launch does not read this variable.

For a single evaluation request, precedence is: explicit CLI or SDK value > environment variable > configuration > default. For a multi-request orchestration, precedence is: explicit CLI or SDK value > orchestration runtime > configuration > default.

Enable auto-install only for trusted components because it modifies the current Python environment. It applies only to paths marked as supported in the table; remote mini_swe_agent, OpenEvolve, and TauBench do not trigger generic auto-install. Local mini_swe_agent also auto-installs only when the package is absent; an installed version mismatch still fails. The generic check confirms only that the current interpreter can discover the declared modules; it does not validate versions or audit the full transitive dependency graph. uv or pip still resolves transitive dependencies during installation.

Check Timing

AgentCompass runs only dependency checks declared by the active evaluation path; it does not scan every extra. A satisfied check completes silently.

Check type	Timing	Location
Benchmark dependency preflight	Before task loading	Current Python interpreter running AgentCompass
Local harness dependency check	Before the harness session starts	Current Python interpreter running AgentCompass
In-Environment component dependency check	Harness session startup or task preparation	Python that executes the component in the selected Environment

Special Cases and Dependency Conflicts

TauBench

Unless you explicitly configure another image, the taubench_docker Recipe selects one with the TauBench dependencies preinstalled, so the Python interpreter running AgentCompass does not need the taubench extra. With host_process, TauBench checks tau2==1.0.1 through python3 in the selected Environment before each task; install both steps below into that same interpreter.

From a source checkout:

TAUBENCH_PYTHON="$(command -v python3)"
uv pip install --python "$TAUBENCH_PYTHON" -e ".[taubench]"
uv pip install --python "$TAUBENCH_PYTHON" --no-deps \
  "tau2 @ git+https://github.com/sierra-research/tau2-bench@v1.0.1"

For an installed release, replace -e ".[taubench]" with "agentcompass[taubench]". Installing the pinned tau2 source separately with --no-deps prevents its transitive dependencies from changing AgentCompass’s shared LiteLLM or Tenacity versions.

The terminal_use, terminal_use_write, alltools, and alltools-qwen retrieval methods for banking_knowledge also need system tools in the execution Environment; the default Docker image already provides them. See TauBench banking_knowledge retrieval configuration to prepare another Environment.

Related Pages

• Install AgentCompass
• Environments
• Common failures