Skip to content

No-LLM Mode

What it does

Set AGENTIC_NO_LLM=1 to run the native and LangChain engines without provider credentials. Both execution paths short-circuit every LLM call and return a deterministic placeholder string: [AGENTIC_NO_LLM placeholder] Set AGENTIC_NO_LLM=0 and a provider key to get real output. Provider probing is skipped entirely — no spurious errors from missing API keys.

How to enable

Bash / PowerShell:

export AGENTIC_NO_LLM=1
agentic run test_deterministic

Windows PowerShell:

$env:AGENTIC_NO_LLM = "1"
agentic run test_deterministic

Expected output: the workflow runs to completion, all steps emit placeholder text, and the process exits with status 0. No provider credentials required.

LangChain engine requires an extra. The native DAG executor works with just the base pip install -e . — that's the common case. If you want to exercise the LangChain adapter under this flag (--engine langchain), install with the extra: pip install -e .[langchain]. Without it, get_chat_model() raises ImportError: langchain-core is required... even when the flag is set. Accepted values for the env var (case-insensitive): 1/true/yes/on are true; empty/0/false/no/off are false; anything else is coerced to false with a logged warning.

The UI also works: run agentic serve + npm run dev under this flag. DAG streaming shows placeholder content at each node.

What works

  • Native DAG executor (both YAML workflow execution and agent-step paths)
  • LangChain adapter (--engine langchain in CLI or config)
  • DAG streaming over WebSocket/SSE (one chunk per message; the entire placeholder is a single chunk)
  • UI demos end-to-end
  • CLI validation commands (agentic list and agentic validate don't invoke an LLM regardless of this flag; agentic compare exercises workflows and honors the flag)
  • RAG pipeline (embeddings already deterministic via SHA-256 hashing)

Scope limits — read this before you report a bug

Not a simulator. The placeholder is a fixed string. Any downstream consumer that parses structured output — PydanticOutputParser, JSON-extraction steps, evaluation rubrics that score against a reference — will fail or score zero. This is by design. Use this mode for shape and flow testing only, not for semantics or behavior validation.

Evaluation runs need real keys. agentic-v2-eval is untouched; scoring rubrics against the placeholder will score zero. Run evaluation workflows with actual provider credentials.

Streaming is not token-by-token. Both complete_stream (native) and astream (LangChain) yield the entire placeholder as one chunk. Useful for testing UI plumbing, not for testing per-token streaming UX.

Tool-calling is a no-op. bind_tools(...) is accepted but ignored. The placeholder is returned regardless of what tools were bound. Workflows that branch on tool calls will always take the no-tool path.

When NOT to use

  • Production deployments
  • Evaluation and rubric-scoring runs
  • Capacity planning or load tests
  • Anything asserting on real model content or behavior

Troubleshooting

I set the var but got "GITHUB_TOKEN environment variable is required"

The get_settings function uses @lru_cache. The flag must be set before the first call to get_settings() in a process. For CLI users, this is automatic (fresh process). For long-running sessions (e.g., interactive notebooks), call:

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

get_settings.cache_clear()
reset_client()

Then set your env var and proceed.

The placeholder isn't showing in the UI

Check your browser's Network tab or console for /api/runs/{run_id}/stream and /ws/execution/{run_id} events. The string [AGENTIC_NO_LLM placeholder] will appear in each node's output field in the DAG.

Can I make it fail instead of returning a placeholder?

No. The flag is for graceful no-op, not for hard errors. If you want to enforce a provider key, unset the flag and let the provider's missing-credential ValueError surface.