Execution Resource Examples

July 31, 2026 ยท View on GitHub

These examples show the recommended developer path for managed execution capabilities.

Most application code should start with agent.enable_* helpers or built-in action packages. They hide the core manager/provider lifecycle and expose model-callable Actions with sensible defaults.

Start Here

FileAudienceWhat it shows
01_action_python_resource_local.pyApp developers, no model requiredEnable a trusted-local Python action with agent.enable_python(..., sandbox="trusted_local") and call it directly for a deterministic local check.
02_agent_python_resource_ollama.pyApp developers with local OllamaLet a model decide to call the enabled Python action before replying.
03_agent_issue_processor_deepseek.pyApp developers with DeepSeekA realistic issue-triage task where the model uses Python for deterministic metrics and then writes the summary.
04_triggerflow_python_resource_local.pyWorkflow/framework developersLet a TriggerFlow chunk invoke the canonical Workspace-bound Python CodeExecution Action.
05_action_nodejs_resource_local.pyApp developers, no model requiredEnable a Node.js CodeExecution action and execute it through the explicit trusted-local provider.
06_action_sqlite_resource_local.pyApp developers, no model requiredEnable a managed SQLite query action against a local database file.
07_browser_resource_browse_local.pyAction/plugin developersBrowse a local page through Browser Execution Resource.
08_health_check_reuse_local.pyProvider/plugin developersShow V2 health-check-before-reuse behavior with a custom provider.
09_action_common_code_runtime_local.pyApp developers, no model requiredRun a Workspace-backed enable_code_runtime(...) Action with ordered providers and an explicit unsafe fallback.

Copy-Paste Shape

For application developers, the shape is intentionally small:

agent = Agently.create_agent()
agent.enable_python(desc="Use for exact calculations. Print the final result as JSON.")

turn = agent.input("Use Python to calculate the average of [15, 23, 42, 8, 12].")
records = agent.get_action_result(prompt=turn.prompt)
result = turn.get_result()

You normally do not need to call Agently.execution_resource directly. The Action dispatcher ensures and releases the managed environment when the enabled action is called.

Example Details

  • 01_action_python_resource_local.py
    • Runs without any model API key.
    • Uses agent.enable_python(..., sandbox="trusted_local"), then calls the registered action directly.
    • Good for verifying the local package without Docker or a model endpoint.
  • 02_agent_python_resource_ollama.py
    • Uses an Ollama OpenAI-compatible endpoint.
    • Defaults to qwen2.5:7b, which is sufficient for the small action-selection task.
    • Lets the model choose the default Docker-backed Python action, then prints action records and the final reply.
  • 03_agent_issue_processor_deepseek.py
    • Uses DeepSeek for a more complex issue-processing prompt.
    • Shows that execution is real: model planning calls the Docker-backed Python action, the sandbox computes metrics, and the final reply uses those action results.
  • 04_triggerflow_python_resource_local.py
    • Runs without any model API key.
    • Keeps TriggerFlow responsible for orchestration and invokes a CodeExecution Action from one chunk.
    • Shows the Workspace materialization and action-call resource boundary without a model API key.
  • 05_action_nodejs_resource_local.py
    • Runs without any model API key.
    • Requires node on PATH; otherwise it prints a skip message.
    • Demonstrates agent.enable_nodejs(..., sandbox="trusted_local") and action-call-scoped release.
  • 06_action_sqlite_resource_local.py
    • Runs without any model API key.
    • Creates a temporary SQLite database and queries it through agent.enable_sqlite(...).
  • 07_browser_resource_browse_local.py
    • Runs without any model API key.
    • Requires Playwright and Chromium; otherwise it prints a skip message.
    • Demonstrates Browse(use_browser_environment=True) with a managed browser resource.
  • 08_health_check_reuse_local.py
    • Runs without any model API key.
    • Creates a local manager and provider to show that unhealthy ready handles are released and replaced before reuse.
  • 09_action_common_code_runtime_local.py
    • Runs without any model API key.
    • Tries an eligible local Docker provider first, then uses the explicitly authorized unsafe local fallback when Python is installed.
    • Demonstrates Workspace grant, ordered provider selection, and the Python/Node.js/Go/C++ adapter catalog.

Before running the Ollama example, make sure Ollama is running and the model is available:

ollama pull qwen2.5:7b

The Ollama and DeepSeek Python-action examples use the default Docker-backed runtime profile. Make sure the local Docker CLI and daemon are available, or pass sandbox="trusted_local" only in trusted compatibility examples. The common-language code runtime example uses isolation="preferred" and unsafe_fallback=True deliberately. Production code that requires isolation should keep the default isolation="required" and omit the unsafe fallback.

Optional Ollama environment variables:

  • OLLAMA_BASE_URL, defaults to http://localhost:11434/v1
  • OLLAMA_DEFAULT_MODEL, defaults to qwen2.5:7b

Before running the DeepSeek example, set:

  • DEEPSEEK_API_KEY
  • DEEPSEEK_BASE_URL, optional, defaults to https://api.deepseek.com/v1
  • DEEPSEEK_DEFAULT_MODEL, optional, defaults to deepseek-v4-flash

Run:

python examples/execution_resource/01_action_python_resource_local.py
python examples/execution_resource/02_agent_python_resource_ollama.py
python examples/execution_resource/03_agent_issue_processor_deepseek.py
python examples/execution_resource/04_triggerflow_python_resource_local.py
python examples/execution_resource/05_action_nodejs_resource_local.py
python examples/execution_resource/06_action_sqlite_resource_local.py
python examples/execution_resource/07_browser_resource_browse_local.py
python examples/execution_resource/08_health_check_reuse_local.py
python examples/execution_resource/09_action_common_code_runtime_local.py

Notes:

  • Execution Resource declarations are lazy; a declaration does not start a sandbox or transport.
  • Business examples should prefer agent.enable_python(...), agent.enable_shell(...), agent.enable_task_workspace_file_actions(...), agent.enable_nodejs(...), agent.enable_code_runtime(...), and agent.enable_sqlite(...) over direct manager/provider APIs. enable_code_runtime(...) is provider-neutral; unsafe local execution requires explicit fallback and isolation configuration.
  • Built-in providers currently cover MCP, Bash, Docker, provider-neutral code execution, Browser, and SQLite. Python, Node.js, Go, and C++ share the code_execution kind and differ only through language adapters. Search is intentionally not an Execution Resource provider; configure proxy, timeout, backend, and region on agently.builtins.actions.Search(...).
  • Ready handles are health-checked before reuse. Unhealthy handles emit execution_resource.unhealthy, are released, and are replaced with fresh handles.
  • enable_* helpers provide default action descriptions, so desc= is optional. By default desc= appends extra guidance; desc_mode="override" replaces the default description only when you need full control.
  • Action dispatch ensures required environments immediately before executor calls.
  • action_call scoped handles are released after the action call.
  • TriggerFlow still exposes non-code live resources through runtime_resources; Workspace-bound code should run through CodeExecution Actions so grants, bundles, provider selection, evidence, and release stay on one path.