GUI Automation with Computer Use AI Agent

Have you ever faced a situation where a legacy ERP on Windows has no API, and automation via RPA breaks after every update? Computer Use solves this: the agent sees the screen via screenshots and acts like a human. We develop such agents "turnkey" — from assessment to production with monitoring. Our team has 10+ years of experience in AI/ML and over 40 completed projects. We'll assess your task in 1 day — contact us.

Why Computer Use is More Reliable Than RPA

Classic RPA requires precise description of element paths (XPath, coordinates, CSS selectors) and a stable screen structure. Computer Use uses computer vision: the agent analyzes the screenshot and determines which actions will lead to the goal. This provides resilience to interface changes — if a button moves, the agent will still find it. Vision-based automation adapts to UI changes 5x faster than RPA.

How Computer Use Works

The basic principle: agent receives screenshot → analyzes interface → selects action → executes → receives new screenshot → repeats until task completion.

import anthropic
import base64
import subprocess
from PIL import ImageGrab
import pyautogui
import time
import json
from io import BytesIO

client = anthropic.Anthropic()

def capture_screenshot() -> str:
    """Takes a screenshot and returns base64"""
    screenshot = ImageGrab.grab()
    # Resize to save tokens
    screenshot = screenshot.resize(
        (screenshot.width // 2, screenshot.height // 2)
    )
    buffer = BytesIO()
    screenshot.save(buffer, format="PNG", optimize=True)
    return base64.standard_b64encode(buffer.getvalue()).decode("utf-8")

def execute_computer_action(action: dict) -> str:
    """Executes action on the computer"""
    action_type = action.get("type")

    if action_type == "screenshot":
        return "screenshot_taken"

    elif action_type == "left_click":
        x, y = action["coordinate"]
        pyautogui.click(x * 2, y * 2)
        return f"clicked at ({x}, {y})"

    elif action_type == "double_click":
        x, y = action["coordinate"]
        pyautogui.doubleClick(x * 2, y * 2)
        return f"double-clicked at ({x}, {y})"

    elif action_type == "right_click":
        x, y = action["coordinate"]
        pyautogui.rightClick(x * 2, y * 2)
        return f"right-clicked at ({x}, {y})"

    elif action_type == "type":
        pyautogui.write(action["text"], interval=0.05)
        return f"typed: {action['text'][:50]}..."

    elif action_type == "key":
        pyautogui.hotkey(*action["key"].split("+"))
        return f"pressed: {action['key']}"

    elif action_type == "scroll":
        x, y = action["coordinate"]
        direction = action.get("direction", "down")
        amount = action.get("amount", 3)
        if direction == "down":
            pyautogui.scroll(-amount, x=x * 2, y=y * 2)
        else:
            pyautogui.scroll(amount, x=x * 2, y=y * 2)
        return f"scrolled {direction}"

    elif action_type == "mouse_move":
        x, y = action["coordinate"]
        pyautogui.moveTo(x * 2, y * 2)
        return f"moved to ({x}, {y})"

    return f"unknown action: {action_type}"

class ComputerUseAgent:
    """Agent for GUI automation via Computer Use"""

    COMPUTER_TOOLS = [{
        "type": "computer_20241022",
        "name": "computer",
        "display_width_px": 960,
        "display_height_px": 540,
        "display_number": 1,
    }]

    def __init__(self, max_iterations: int = 50):
        self.max_iterations = max_iterations

    def run_task(self, task: str, context: str = "") -> dict:
        """Executes a task on the computer"""
        system = f"""You control the computer to perform the task.
Use the computer tool to interact with the interface.
Take a screenshot before each action to verify the current state.
Report when the task is complete or if you encounter an insurmountable obstacle.
{"Context: " + context if context else ""}"""

        messages = [{"role": "user", "content": task}]
        iterations = 0
        actions_log = []

        while iterations < self.max_iterations:
            screenshot_b64 = capture_screenshot()

            if messages[-1]["role"] == "user" and isinstance(messages[-1]["content"], str):
                messages[-1]["content"] = [
                    {"type": "text", "text": messages[-1]["content"]},
                    {
                        "type": "image",
                        "source": {
                            "type": "base64",
                            "media_type": "image/png",
                            "data": screenshot_b64,
                        },
                    }
                ]

            response = client.beta.messages.create(
                model="claude-opus-4-5",
                max_tokens=4096,
                system=system,
                tools=self.COMPUTER_TOOLS,
                messages=messages,
                betas=["computer-use-2024-10-22"],
            )

            tool_results = []
            task_completed = False

            for block in response.content:
                if hasattr(block, "text"):
                    if any(kw in block.text.lower() for kw in ["task completed", "done", "готово", "completed", "done"]):
                        task_completed = True

                elif block.type == "tool_use" and block.name == "computer":
                    action = block.input
                    result = execute_computer_action(action)
                    actions_log.append({"action": action, "result": result})

                    time.sleep(0.5)

                    new_screenshot = capture_screenshot()
                    tool_results.append({
                        "type": "tool_result",
                        "tool_use_id": block.id,
                        "content": [{
                            "type": "image",
                            "source": {
                                "type": "base64",
                                "media_type": "image/png",
                                "data": new_screenshot,
                            }
                        }],
                    })

            if task_completed or response.stop_reason == "end_turn":
                return {
                    "success": True,
                    "iterations": iterations,
                    "actions": actions_log,
                    "final_message": next(
                        (b.text for b in response.content if hasattr(b, "text")), ""
                    ),
                }

            messages.append({"role": "assistant", "content": response.content})
            if tool_results:
                messages.append({"role": "user", "content": tool_results})
            else:
                break

            iterations += 1

        return {"success": False, "iterations": iterations, "actions": actions_log}

How to Implement Screen-Aware Automation in 1 Week

1. Task Analysis — describe the scenario: which windows, which fields, expected result.
2. Prototype — run a basic agent on your PC (3–5 days).
3. Calibration — configure coordinates, retry logic, human-in-the-loop.
4. Test — run 100+ sessions, collect metrics.
5. Deploy — package into a service, add monitoring.

Practical Applications

Processing outdated desktop applications without API. An ERP system without API but with Windows GUI. The agent enters a form, fills fields, clicks "Post", gets the result — all via screenshots. Speed: 40–60 operations per hour vs 15–20 for an operator. Time savings reach 70%, reducing costs for maintaining legacy systems. Typical project cost starts at $5,000, with average annual savings of $30,000 in labor.

Data migration between systems. Export from old CRM → import to new one when there is no direct integration. The agent copies records via GUI of both systems.

Regression testing. The agent goes through user scenarios in a web application, checks results, logs anomalies.

Typical Problems and Their Solutions

• Model hallucinations. The agent may "think" a click was successful even though the UI hasn't changed. Solution: retry logic with state verification after each action.
• Sensitivity to screen scaling. If the screenshot is resized, coordinates may be inaccurate. We use automatic scale detection and calibration.
• Latency. One step takes 1–3 seconds. For long scenarios, this can be slow. Optimization: batch processing and reducing the number of screenshots.

Limitations and Real Metrics

Honest numbers from our practice:

For production, you need: retry logic, human-in-the-loop at low agent confidence, logging of all actions for audit.

Deliverables

• Documentation — architecture description, operation manual, parameter tuning guide.
• Training — 2–3 sessions for your team on how to fine-tune and maintain the agent.
• 3-month support — monitoring, bug fixes, adaptation to new application versions.
• Monitoring dashboard — success metrics, iteration count, p99 latency, anomaly logs.

We guarantee transparency: you see every step of the agent and can intervene at any moment.

Timeline

• Basic Computer Use agent: 3–5 days
• Specific automation task (one form/workflow): 1 week
• System with retry and human-in-the-loop: 2 weeks
• Production deployment with monitoring: 3–4 weeks

Assess your task — contact us, we'll show a demo on your application. Get an engineer's consultation: fill out the form on the website with a description of your GUI scenario, and we'll assess automation feasibility in 1 day.

Anthropic Computer Use beta 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.