AI Agent Orchestrator Development (Agent Orchestration)

Imagine you have five AI agents. Each has its own expertise — analyzing finances, checking legal documents, assessing the market. You command a company due diligence. Without an orchestrator, chaos begins: agents are called sequentially, errors restart everything, parallel tasks are missing, and p99 latency exceeds a minute. We encountered this on a real project and built an orchestrator with LangGraph. It cut DD execution time from 4 weeks to 3 days. The orchestrator is 9 times faster. Aspect coverage reaches 94%. This is not the limit — even greater efficiency is possible.

Why are agents ineffective without an orchestrator?

A typical problem: each agent works in isolation. You pass a request from researcher to analyst to writer. An error at any stage restarts the entire chain. If one agent hangs, the whole process stalls. Parallel tasks are absent. The result is low GPU utilization (about 15%), huge latency, and constant manual fixes. Manual due diligence takes up to 4 weeks. For a large project, costs can exceed $100,000.

How does LangGraph simplify orchestration?

LangGraph is a graph-based framework. It describes dependencies between agents as a directed graph. Nodes are agents. Edges are conditional transitions. Built-in checkpoints via MemorySaver save state after each step. Retry logic and fallback are out-of-the-box. Below is an implementation of an orchestrator with LangGraph that we use in commercial projects.

from langgraph.graph import StateGraph, END
from langgraph.checkpoint.memory import MemorySaver
from typing import TypedDict, Annotated
from langchain_openai import ChatOpenAI
import operator
import json

class OrchestratorState(TypedDict):
    user_request: str
    task_plan: list[dict]           # [{task_id, description, agent, status, result}]
    current_task_index: int
    agent_results: Annotated[dict, lambda a, b: {**a, **b}]
    final_response: str
    error_count: int

llm = ChatOpenAI(model="gpt-4o", temperature=0)

# Agent registry
AGENT_REGISTRY = {
    "researcher": ResearcherAgent(),
    "analyst": AnalystAgent(),
    "writer": WriterAgent(),
    "sql_agent": SQLAgent(),
    "code_interpreter": CodeInterpreterAgent(),
    "file_processor": FileProcessorAgent(),
}

def plan_tasks(state: OrchestratorState) -> OrchestratorState:
    """Orchestrator breaks down task into subtasks and assigns agents"""
    available_agents = list(AGENT_REGISTRY.keys())
    response = llm.invoke(f"""
Break down the following task into subtasks and assign an agent for each.
Available agents: {available_agents}

Task: {state["user_request"]}

Return a JSON list:
[{{"task_id": "t1", "description": "...", "agent": "researcher", "dependencies": []}}]
Dependencies: list of task_ids that must be completed before this task.
""")
    task_plan = json.loads(response.content)
    for task in task_plan:
        task["status"] = "pending"
        task["result"] = None
    return {**state, "task_plan": task_plan, "current_task_index": 0}

def execute_next_task(state: OrchestratorState) -> OrchestratorState:
    """Executes the next ready task"""
    task_plan = state["task_plan"].copy()
    # Find the next task whose all dependencies are completed
    next_task = None
    for task in task_plan:
        if task["status"] == "pending":
            deps_completed = all(
                any(t["task_id"] == dep and t["status"] == "completed"
                    for t in task_plan)
                for dep in task.get("dependencies", [])
            )
            if deps_completed:
                next_task = task
                break
    if not next_task:
        return {**state, "current_task_index": -1}  # All tasks completed
    # Execute task via corresponding agent
    agent = AGENT_REGISTRY.get(next_task["agent"])
    if not agent:
        next_task["status"] = "failed"
        next_task["result"] = f"Agent {next_task['agent']} not found"
    else:
        dependency_results = {
            dep: state["agent_results"].get(dep)
            for dep in next_task.get("dependencies", [])
        }
        try:
            result = agent.execute(
                task=next_task["description"],
                context=dependency_results,
            )
            next_task["status"] = "completed"
            next_task["result"] = result
        except Exception as e:
            next_task["status"] = "failed"
            next_task["result"] = str(e)
    # Update plan
    updated_plan = [
        task if task["task_id"] != next_task["task_id"] else next_task
        for task in task_plan
    ]
    return {
        **state,
        "task_plan": updated_plan,
        "agent_results": {next_task["task_id"]: next_task["result"]},
    }

def should_continue(state: OrchestratorState) -> str:
    """Determines the next step of the orchestrator"""
    pending = [t for t in state["task_plan"] if t["status"] == "pending"]
    failed = [t for t in state["task_plan"] if t["status"] == "failed"]
    if failed and state["error_count"] >= 3:
        return "finalize_with_errors"
    if not pending:
        return "aggregate_results"
    return "execute_next"

def aggregate_results(state: OrchestratorState) -> OrchestratorState:
    """Aggregates all agent results into a final response"""
    all_results = {t["task_id"]: t["result"] for t in state["task_plan"]}
    final = llm.invoke(f"""
Based on the results from different agents, form a final response.
Original request: {state["user_request"]}
Results: {json.dumps(all_results, ensure_ascii=False)}
""").content
    return {**state, "final_response": final}

# Build graph
graph = StateGraph(OrchestratorState)
graph.add_node("plan", plan_tasks)
graph.add_node("execute_next", execute_next_task)
graph.add_node("aggregate_results", aggregate_results)
graph.set_entry_point("plan")
graph.add_edge("plan", "execute_next")
graph.add_conditional_edges("execute_next", should_continue, {
    "execute_next": "execute_next",
    "aggregate_results": "aggregate_results",
    "finalize_with_errors": "aggregate_results",
})
graph.add_edge("aggregate_results", END)
orchestrator = graph.compile(checkpointer=MemorySaver())

How does the orchestrator execute tasks in parallel?

import asyncio

async def execute_parallel_tasks(tasks_batch: list[dict]) -> list[dict]:
    """Parallel execution of independent tasks"""
    coroutines = []
    for task in tasks_batch:
        agent = AGENT_REGISTRY.get(task["agent"])
        if agent:
            coroutines.append(asyncio.to_thread(agent.execute, task=task["description"]))
    results = await asyncio.gather(*coroutines, return_exceptions=True)
    for task, result in zip(tasks_batch, results):
        if isinstance(result, Exception):
            task["status"] = "failed"
            task["result"] = str(result)
        else:
            task["status"] = "completed"
            task["result"] = result
    return tasks_batch

Practical case: orchestrator for due diligence

From our practice: automated company vetting for M&A. Parallel work of 5 agents:

1. Financial Agent: analysis of 3 years of reporting
2. Legal Agent: checking litigation, restrictions
3. HR Agent: personnel structure, turnover
4. Market Agent: market position, competitors
5. Risk Agent: synthesis of risks from all sources

Execution graph:

• t1 (financial), t2 (legal), t3 (hr), t4 (market) — parallel
• t5 (risk) — depends on t1, t2, t3, t4
• t6 (final_report) — depends on t5

Results:

• DD time: 4 weeks → 3 days (9x faster)
• Aspect coverage: 78% → 94%
• Cost per DD reduced by 71% — savings amount to tens of thousands of dollars per project. For large corporations, savings can reach hundreds of thousands. For example, one client saved $70,000 annually.

Let's compare approaches:

Task distribution among agents

Monitoring and tracing

import mlflow

def log_orchestration_run(state: OrchestratorState):
    with mlflow.start_run():
        mlflow.log_metrics({
            "total_tasks": len(state["task_plan"]),
            "completed_tasks": sum(1 for t in state["task_plan"] if t["status"] == "completed"),
            "failed_tasks": sum(1 for t in state["task_plan"] if t["status"] == "failed"),
        })
        mlflow.log_text(json.dumps(state["task_plan"], indent=2), "task_execution_log.json")

Scope of work

• Designing a multi-agent system architecture for your task
• Implementing the orchestrator using LangGraph or a comparable framework
• Developing specialized AI agents (up to 10) integrated with your data
• Configuring parallel execution, retry logic, and fallback
• Monitoring and tracing via MLflow, Weights & Biases
• Documentation: graph schema, agent APIs, operation manual
• Training your team (2–3 sessions)

Estimated timeline

• Orchestrator design: 1–2 weeks
• Implementation of base agents (3–5): 3–5 weeks
• Parallel execution integration: 1 week
• Error handling and monitoring: 1–2 weeks
• Total: 6–10 weeks

We guarantee stable production operation. All solutions are covered by unit tests and integration tests. We have 10+ years of experience in AI/ML and have implemented orchestrators for financial due diligence, report automation, and content generation.

We offer turnkey AI agent orchestrator development in 6–10 weeks. Write to us for a free project assessment. Contact us — we will evaluate your project, select architecture, and determine implementation timelines.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.