AI Workflow Orchestration: Chains, DAGs, Human-in-the-Loop & Production Patterns

A single LLM call is rarely enough. Real AI applications chain multiple calls, branch on conditions, wait for human approval, and recover from failures. This is workflow orchestration — the layer between your LLM and production.

Why Orchestration Matters#

A simple chatbot is one LLM call. A production AI system looks like this:

User submits insurance claim
  → Extract claim details (LLM)
  → Classify severity (LLM)
  → IF high severity → route to human reviewer
  → IF low severity → auto-approve
  → Generate response letter (LLM)
  → Send notification
  → Log to audit trail

Each step can fail, each branch has different logic, and some steps need human oversight. You need orchestration.

LLM Chains#

The simplest pattern: feed one LLM's output into the next.

# Sequential chain: summarize → translate → format
summary = llm.invoke("Summarize this document: {doc}")
translation = llm.invoke(f"Translate to Spanish: {summary}")
formatted = llm.invoke(f"Format as a professional email: {translation}")

Chains are linear — step 2 always follows step 1. They're easy to understand but can't express branches or parallel paths.

Chain Limitations#

• No conditional logic (every step always runs)
• No parallel execution
• One failure breaks the entire chain
• Can't pause for human input

DAG-Based Workflows#

Directed Acyclic Graphs (DAGs) solve these limitations. Each node is a step, edges define data flow, and the graph can branch and merge:

                    ┌──────────────┐
                    │  Extract     │
                    │  entities    │
                    └──────┬───────┘
                           │
                    ┌──────▼───────┐
                    │  Classify    │
                    │  intent      │
                    └──────┬───────┘
                      ╱          ╲
               ┌─────▼────┐  ┌───▼──────────┐
               │ Auto-     │  │ Human        │
               │ respond   │  │ review queue │
               └─────┬─────┘  └───┬──────────┘
                      ╲          ╱
                    ┌──▼──────▼──┐
                    │  Send      │
                    │  response  │
                    └────────────┘

DAGs enable parallelism (independent nodes run concurrently), branching (different paths based on conditions), and merging (collecting results from parallel branches).

Conditional Branching#

Most AI workflows need to make decisions mid-flow:

def route_claim(state):
    severity = state["severity"]
    if severity == "high":
        return "human_review"
    elif severity == "medium":
        return "senior_agent"
    else:
        return "auto_approve"

Conditions can be based on:

• LLM classification output (sentiment, category, severity)
• Confidence scores (low confidence routes to human)
• Business rules (claims over $10K always need review)
• External data (customer tier, account age)

Human-in-the-Loop#

Some decisions are too important for full automation. Human-in-the-loop (HITL) pauses the workflow and waits for a person:

AI classifies medical image → 95% confidence "benign"
  → Auto-approve? No. Route to radiologist.

AI drafts legal contract → Looks good
  → Auto-send? No. Lawyer reviews first.

Implementation Patterns#

Approval gates:

# Workflow pauses here until human approves
async def human_review_node(state):
    # Save state to database
    await save_pending_review(state)
    # Webhook/polling resumes workflow when human approves
    approval = await wait_for_human_approval(state["review_id"])
    state["approved"] = approval.approved
    state["reviewer_notes"] = approval.notes
    return state

Confidence thresholds:

def should_escalate(state):
    if state["confidence"] < 0.85:
        return "human_review"
    return "auto_process"

Edit-and-continue: The AI generates a draft, human edits it, and the workflow continues with the edited version. Common for content generation, email drafting, and report writing.

Error Handling#

Workflows fail. Models hallucinate. APIs time out. Robust orchestration handles all of it.

Retry with Exponential Backoff#

@retry(max_attempts=3, backoff=exponential(base=2))
async def call_llm(prompt):
    return await llm.invoke(prompt)

Fallback Chains#

def summarize_with_fallback(document):
    try:
        return call_model("claude-sonnet", f"Summarize: {document}")
    except ModelOverloaded:
        return call_model("gpt-4o-mini", f"Summarize: {document}")
    except Exception:
        return extractive_summary(document)  # Non-LLM fallback

Output Validation#

def validate_extraction(state):
    result = state["extracted_data"]
    if not result.get("name") or not result.get("email"):
        state["retry_count"] = state.get("retry_count", 0) + 1
        if state["retry_count"] < 3:
            return "retry_extraction"
        return "human_review"
    return "next_step"

Dead Letter Queues#

Failed workflows go to a dead letter queue for manual inspection rather than being silently dropped.

LangChain#

LangChain provides building blocks for LLM workflows:

from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langchain_core.output_parsers import StrOutputParser

prompt = ChatPromptTemplate.from_template("Summarize: {text}")
model = ChatOpenAI(model="gpt-4o")
parser = StrOutputParser()

# LCEL chain — composable with the | operator
chain = prompt | model | parser
result = chain.invoke({"text": "Long document here..."})

Strengths: Large ecosystem, many integrations, good for prototyping. Limitations: Abstraction overhead, debugging can be opaque, less control over execution flow.

LangGraph#

LangGraph extends LangChain with stateful, graph-based workflows:

from langgraph.graph import StateGraph, END

# Define state schema
class WorkflowState(TypedDict):
    query: str
    classification: str
    response: str

# Build graph
graph = StateGraph(WorkflowState)

graph.add_node("classify", classify_query)
graph.add_node("handle_billing", handle_billing)
graph.add_node("handle_technical", handle_technical)
graph.add_node("respond", generate_response)

graph.set_entry_point("classify")
graph.add_conditional_edges("classify", route_by_class, {
    "billing": "handle_billing",
    "technical": "handle_technical"
})
graph.add_edge("handle_billing", "respond")
graph.add_edge("handle_technical", "respond")
graph.add_edge("respond", END)

app = graph.compile()
result = app.invoke({"query": "Why was I charged twice?"})

LangGraph supports persistence (save and resume workflows), human-in-the-loop (interrupt nodes), and streaming (token-by-token output).

Temporal for AI Workflows#

Temporal is a durable execution platform — workflows survive process crashes, server restarts, and network failures:

@workflow.defn
class ClaimProcessingWorkflow:
    @workflow.run
    async def run(self, claim_data: dict):
        # Each activity is automatically retried on failure
        extracted = await workflow.execute_activity(
            extract_claim_details, claim_data,
            start_to_close_timeout=timedelta(seconds=30)
        )

        severity = await workflow.execute_activity(
            classify_severity, extracted,
            start_to_close_timeout=timedelta(seconds=15)
        )

        if severity == "high":
            # Wait up to 48 hours for human approval
            approved = await workflow.wait_condition(
                lambda: self.approval_received,
                timeout=timedelta(hours=48)
            )
            if not approved:
                return {"status": "escalated_to_manager"}

        response = await workflow.execute_activity(
            generate_response, extracted,
            start_to_close_timeout=timedelta(seconds=30)
        )
        return {"status": "completed", "response": response}

Why Temporal for AI: Durable timers (wait days for human input), automatic retries, full execution history, and replay debugging.

Prefect for AI Pipelines#

Prefect excels at data-oriented AI workflows — batch processing, ETL with LLM enrichment, scheduled jobs:

from prefect import flow, task

@task(retries=3, retry_delay_seconds=10)
def embed_documents(documents):
    return embedding_model.encode(documents)

@task
def store_in_vector_db(embeddings, documents):
    vector_store.upsert(embeddings, documents)

@flow(name="daily-embedding-pipeline")
def embedding_pipeline():
    new_docs = fetch_new_documents()
    embeddings = embed_documents(new_docs)
    store_in_vector_db(embeddings, new_docs)
    return {"processed": len(new_docs)}

# Schedule to run daily
embedding_pipeline()

Prefect provides observability dashboards, scheduling, and infrastructure management out of the box.

Choosing Your Orchestration Tool#

Many production systems combine tools — LangGraph for the agent logic, Temporal for durability, Prefect for batch jobs.

Key Takeaways#

1. Chains are linear — fine for simple pipelines, insufficient for real workflows
2. DAGs enable branching and parallelism — model your workflow as a graph
3. Human-in-the-loop is not optional for high-stakes decisions
4. Error handling must be a first-class concern, not an afterthought
5. Pick the right tool — LangGraph for agents, Temporal for durability, Prefect for batch
6. Start simple — plain Python chains, then add orchestration when complexity demands it

---

Explore more system design guides at codelit.io/blog.