AI Agent API Integration: Architecture, Security, Case

Your AI agent is trained to answer questions but cannot update a lead status in CRM or verify a counterparty via API? Then its business value is minimal. We solve this by creating agents with secure access to external systems. This article covers architecture, security, and a practical case. Below is code you can adapt to your stack.

Without access to external services, the agent operates in isolation. Only with up-to-date data from CRM, ERP, or payment systems can the agent perform business tasks: update deals, verify counterparties, send notifications. Our engineers have years of experience in MLOps and API integrations; we have delivered 20+ projects for CRM and ERP systems. We guarantee your agent will work reliably even under peak loads.

Main challenges are authentication, rate limiting, error handling, and security. Without proper architecture, the agent may perform unwanted operations or crash at the first failure. We use retry patterns, permission models, and logging to minimize risks.

What tasks does an AI agent with API solve?

• Real-time updating of lead and deal statuses in CRM.
• Counterparty verification by TIN via FNS API or Dadata.
• Automatic task creation and sending personalized emails.
• Data synchronization between ERP and accounting systems.

Why is an AI agent with API access a must-have for business?

Manual lead processing, constant counterparty checks, and CRM updates consume hours of managers' time. An AI agent with API takes over these tasks. Result: response time drops from 47 minutes to 4 minutes, error count falls by 90%. Average savings per lead is about $7. Payback period for such an agent is less than two months.

How do we design the integration architecture?

The basic pattern is asynchronous HTTP requests with wrappers for retry, auth, and logging. Here's an example class for CRM integration:

from typing import Any, Optional
import httpx
import asyncio
from pydantic import BaseModel

class APITool:
    """Base class for API integration"""

    def __init__(self, base_url: str, api_key: str = None, timeout: int = 10):
        self.base_url = base_url
        self.headers = {"Authorization": f"Bearer {api_key}"} if api_key else {}
        self.timeout = timeout

    async def request(self, method: str, endpoint: str, **kwargs) -> dict:
        async with httpx.AsyncClient(timeout=self.timeout) as client:
            response = await client.request(
                method,
                f"{self.base_url}{endpoint}",
                headers=self.headers,
                **kwargs
            )
            response.raise_for_status()
            return response.json()

# Concrete implementation for CRM
class CRMAPITool(APITool):
    async def get_customer(self, customer_id: str) -> dict:
        return await self.request("GET", f"/customers/{customer_id}")

    async def update_customer_status(self, customer_id: str, status: str) -> dict:
        return await self.request("PATCH", f"/customers/{customer_id}",
                                  json={"status": status})

    async def create_deal(self, customer_id: str, amount: float, stage: str) -> dict:
        return await self.request("POST", "/deals",
                                  json={"customer_id": customer_id, "amount": amount, "stage": stage})

How to ensure API call security?

Direct agent access to API requires guardrails — without them, the agent may perform unwanted operations. We use a combination of permission model and logging:

from functools import wraps
import logging

logger = logging.getLogger(__name__)

class APIPermissionError(Exception):
    pass

# Decorator for permission control
def require_permission(permission: str):
    def decorator(func):
        @wraps(func)
        async def wrapper(*args, permission_context=None, **kwargs):
            if permission_context and not permission_context.has_permission(permission):
                raise APIPermissionError(f"Permission denied: {permission}")
            return await func(*args, **kwargs)
        return wrapper
    return decorator

# Logging all agent API calls
def log_api_call(func):
    @wraps(func)
    async def wrapper(*args, **kwargs):
        logger.info(f"API call: {func.__name__}, args={kwargs}")
        result = await func(*args, **kwargs)
        logger.info(f"API result: {func.__name__} returned {type(result).__name__}")
        return result
    return wrapper

class SafeCRMTool(CRMAPITool):
    @log_api_call
    @require_permission("crm:read")
    async def get_customer(self, customer_id: str) -> dict:
        return await super().get_customer(customer_id)

    @log_api_call
    @require_permission("crm:write")
    async def update_customer_status(self, customer_id: str, status: str) -> dict:
        # Additional validation: allowed statuses
        allowed_statuses = ["active", "inactive", "pending"]
        if status not in allowed_statuses:
            raise ValueError(f"Status must be one of {allowed_statuses}")
        return await super().update_customer_status(customer_id, status)

Error handling for API

import asyncio
from tenacity import retry, stop_after_attempt, wait_exponential, retry_if_exception_type

class APIError(Exception):
    pass

class RateLimitError(APIError):
    pass

@retry(
    stop=stop_after_attempt(3),
    wait=wait_exponential(multiplier=1, min=1, max=10),
    retry=retry_if_exception_type(RateLimitError),
)
async def api_call_with_retry(tool, method, *args, **kwargs):
    try:
        return await getattr(tool, method)(*args, **kwargs)
    except httpx.HTTPStatusError as e:
        if e.response.status_code == 429:
            raise RateLimitError("Rate limit exceeded")
        elif e.response.status_code >= 500:
            raise APIError(f"Server error: {e.response.status_code}")
        raise

# Wrapper function for the agent
def create_api_tool_for_agent(tool_instance, method_name: str) -> callable:
    """Creates a synchronous wrapper for the agent loop"""
    async def async_call(**kwargs) -> str:
        try:
            result = await api_call_with_retry(tool_instance, method_name, **kwargs)
            return json.dumps(result, ensure_ascii=False)
        except APIError as e:
            return json.dumps({"error": str(e), "retry": "automatic"})
        except Exception as e:
            return json.dumps({"error": f"Unexpected: {str(e)}"})

    def sync_call(**kwargs) -> str:
        return asyncio.run(async_call(**kwargs))

    return sync_call

Real-world case: sales agent with CRM, Dadata, and FNS

One of our clients — a company with a 15-person sales department — deployed an AI agent for processing leads from legal entities. The agent integrates with AmoCRM, Dadata, and FNS API. Scenario: new lead — the agent automatically retrieves data from CRM, gets details via Dadata by TIN, checks the counterparty through FNS, creates a task for the manager, and sends a personalized email.

Metrics before and after:

The agent is 11 times faster and reduced manager workload by 68%. Average savings per lead is about $7, which at a flow of 500 leads per month gives $3,500 savings.

Rate Limiting and Cost Control

from asyncio import Semaphore

class RateLimitedAPITool:
    """API with request rate limiting"""

    def __init__(self, api_tool, max_concurrent: int = 5, requests_per_minute: int = 60):
        self.tool = api_tool
        self.semaphore = Semaphore(max_concurrent)
        self.rpm_limit = requests_per_minute

    async def call(self, method: str, **kwargs) -> dict:
        async with self.semaphore:
            return await getattr(self.tool, method)(**kwargs)

How we develop the integration: step-by-step process

1. API audit: document all endpoints, authentication types, and limits.
2. Permission model design: define which actions the agent is allowed to perform.
3. Wrapper implementation: write classes with retry, logging, and error handling.
4. Testing: simulate scenarios, including errors and limit overruns.
5. Deployment: set up monitoring and alerts.

What's included in turnkey development

Timelines and cost

Timelines depend on the number of APIs and authentication complexity. Approximate:

• Integration development (3–5 APIs): 2–4 weeks
• Agent loop with error handling: 1–2 weeks
• Testing and permission model: 1–2 weeks
• Total: 4–8 weeks

Actual cost is calculated individually after analyzing your stack. Order the development of an AI agent for your business. Contact us for a consultation — we will propose an architecture for your CRM and budget. Our engineers guarantee stable agent operation and timely support.

For more on APIs for agents, see OpenAI API Documentation.

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.