Skip to content

Development without a model provider

Set AGENTIC_NO_LLM=1 to replace the runtime's normal native and LangChain model clients with deterministic placeholder implementations. This prevents those clients from calling a provider.

Use this mode to test workflow structure, scheduling, events, persistence, and UI flow. Do not use it to test output meaning or provider behavior.

Enable it

Bash:

export AGENTIC_NO_LLM=1
agentic run test_deterministic \
  --input agentic-workflows-v2/tests/fixtures/deterministic_input.json

PowerShell:

$env:AGENTIC_NO_LLM = "1"
.\.venv\Scripts\agentic.exe run test_deterministic `
  --input agentic-workflows-v2\tests\fixtures\deterministic_input.json

test_deterministic contains two tier-0 agent steps. In the current native and LangGraph paths those agents execute deterministic echo/count handlers without entering a model-client loop. The workflow finishes with Status: SUCCESS and returns the declared processed_text and step_count outputs, making it the preferred adapter smoke test.

Tier-1 and higher agents still use placeholder clients in this mode. Depending on the client path, placeholder output may be a fixed text value or a small mapping such as:

{"placeholder": true, "reason": "llm_unavailable"}

What it covers

  • the shared native LLMClientWrapper;
  • LangChain chat-model construction;
  • normal CLI and API paths that use those clients;
  • WebSocket and server-sent event plumbing for placeholder responses; and
  • deterministic orchestration fallback where implemented.

The LangChain adapter still requires the langchain extra because its graph and message types must be importable:

python -m pip install -e "./agentic-workflows-v2[langchain]"

What it does not prove

Placeholder mode does not test:

  • provider credentials, permissions, availability, or quotas;
  • model routing against live provider catalogs;
  • structured JSON or Pydantic output from a model;
  • tool calls selected by a model;
  • token-by-token streaming;
  • content quality or evaluation quality;
  • latency, retry, fallback, or cost behavior; or
  • provider SDK request and response compatibility.

Placeholder output is deterministic for the selected path, but its shape is not a substitute for a real provider response. A step that requires a specific model schema may fail honestly or take a fallback path. That is not evidence that the same step will fail with a configured model.

Independent features

AGENTIC_NO_LLM does not turn every AI-related component into a fake:

  • agentic-v2-eval can run deterministic rubric and metric scoring without a key, but its LLM-backed evaluators still need an injected live or fake client.
  • Direct provider SDK examples or integrations may have their own credential gates. Confirm that a path uses the shared runtime client before assuming this flag covers it.

Dashboard use

Set the variable in the shell that starts the backend:

$env:AGENTIC_NO_LLM = "1"
just dev

The combined development environment uses FastAPI port 8010 and Vite port 5173. A placeholder response is emitted as one chunk, so this mode tests event delivery but not gradual token streaming.

Disable it

Start a new process without the variable. In PowerShell:

Remove-Item Env:AGENTIC_NO_LLM

In Bash:

unset AGENTIC_NO_LLM

Long-running Python processes cache settings and the shared model client. If a test or notebook deliberately changes the flag after imports, reset both:

from agentic_v2.models.client import reset_client
from agentic_v2.settings import get_settings

get_settings.cache_clear()
reset_client()

Normal applications should restart instead of changing mode inside a running process.

Accepted values

The settings parser treats 1, true, yes, and on as true, ignoring case. It treats 0, false, no, off, and an empty value as false.

When not to use it

  • production or shared deployments;
  • live provider integration tests;
  • quality or correctness claims based on model content;
  • capacity or performance tests; or
  • tool-selection and structured-output tests.