How to Build an AI-Powered Multi-Agent System for Mobile Apps

We’ve seen projects suffer from a single general‑purpose agent. The context window fills up. The model gets confused about which tool to call. Instead of solving the task, it throws errors. In one project, a user asked: 'find a hotel and book it'. The agent started searching for flights because flight search tools were in the same set. A multi‑agent system fixes this. Each agent handles its own domain. An orchestrator coordinates them. Over 5 years of hands‑on mobile AI development (30+ deployments), we’ve built an architecture that reduces hallucinations by 60% and speeds up complex tasks by 2–3 times. Below are patterns, contracts, and real‑world examples from our practice.

Want to see how this works in your scenario? Contact us for a demo.

Why a Single Agent Falls Short

According to Anthropic research, agents with more than 5 tools lose accuracy by 40%. A multi‑agent system breaks down the task:

• Orchestrator — receives the user task, decomposes into subtasks, delegates to specialized agents.
• Research Agent — searches and gathers information (web search, RAG, database).
• Action Agent — executes actions (API calls, bookings).
• Critic Agent — verifies correctness and safety of results.

Classic use case for an AI mobile application: a trip planning agent. Orchestrator gets 'organize a business trip to Warsaw for 3 days'. Research Agent searches flights and hotels. Action Agent books. Critic Agent checks date correctness and price. Orchestrator compiles the final plan.

Choosing a Topology for Mobile Apps

For mobile products, Supervisor with 2‑3 specialized agents on the backend is sufficient. The orchestrator knows each agent’s contract. It does not rely on LLM 'understanding'. Average task completion time is reduced by 40%.

Inter‑Agent Communication: What to Pass

Agents communicate via structured messages, not raw text. Here’s why it matters: if the Research Agent returns unstructured text, the Action Agent may misinterpret. Use JSON contracts:

{
  "agent": "research",
  "task_id": "trip-warsaw",
  "status": "completed",
  "result": {
    "flights": [
      {"id": "LOT123", "price": 189, "departure": "next Monday 06:30"}
    ],
    "hotels": [
      {"id": "H456", "name": "Marriott Warsaw", "price_per_night": 95}
    ]
  }
}

The orchestrator knows each agent’s contract. It does not rely on LLM 'understanding'.

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "type": "object",
  "properties": {
    "agent": {"type": "string"},
    "task_id": {"type": "string"},
    "status": {"type": "string", "enum": ["in_progress", "completed", "failed"]},
    "result": {"type": "object"},
    "error": {"type": "string"}
  },
  "required": ["agent", "task_id", "status"]
}

State Management on the Mobile Client

A multi‑agent process can take 30‑120 seconds. The mobile UI must:

1. Show the current active agent and its step.
2. Allow cancellation at any point.
3. Continue working when the app is backgrounded (push on completion).
4. On agent failure, show partial results.

On Android: WorkManager for background orchestration + StateFlow for UI updates. On iOS: BackgroundTasks framework + AsyncStream.

WebSocket or Server‑Sent Events for real‑time step updates are better than long polling. The client subscribes to a task_id and receives events:

event: agent_step
data: {"agent": "research", "step": "Searching flights Minsk→Warsaw", "progress": 0.3}

event: agent_step
data: {"agent": "action", "step": "Booking flight LOT123", "progress": 0.7}

event: task_complete
data: {"task_id": "trip-warsaw", "result": {...}}

Agent Context Isolation

Each agent should have its own minimal context. Only what is needed for its task. Do not pass booking tool information to the Research Agent, and vice versa. Smaller context means fewer hallucinations and cheaper calls. This is LLM context isolation.

Critically, the Critic Agent receives only the final result. It checks it against a checklist: dates are valid, total matches selected options, no contradictions. This is the last barrier before showing to the user.

Cost and Optimization

A multi‑agent system multiplies LLM calls. To optimize:

• Specialized agents use cheaper models (GPT-4o-mini, Claude Haiku) for routine tasks.
• Orchestrator and Critic use more powerful models (GPT-4o, Claude Sonnet).
• Cache Research Agent results for similar repeated queries (semantic caching).

Savings on LLM calls can reach 40%. Total cost of ownership decreases by 30% due to caching and cheaper models. Implementation cost ranges from $10,000 to $30,000.

What’s Included (Deliverables)

We provide:

• Architecture documentation (topology, contracts, flow diagrams).
• Implemented agents and orchestrator (source code, configuration).
• WebSocket protocol and mobile client integration.
• Progress UI components with SwiftUI Combine and Jetpack Compose.
• Failure and partial result testing.
• Client team training.
• Launch phase support.

Phases and Timelines

1. Analysis and topology design (1–2 weeks).
2. Agent and orchestrator implementation (2–3 weeks).
3. Server orchestrator integration (1–2 weeks).
4. WebSocket protocol for client (1 week).
5. Mobile progress UI (1–2 weeks).
6. Testing and bug fixes (1–2 weeks).

A multi‑agent system with 3 agents and mobile UI — 6–10 weeks turnkey. We’ll assess your project for free — just reach out.

Order a multi‑agent system implementation for your mobile app. Our engineers guarantee stable architecture and help with optimization. Over 5 years of experience and 30+ mobile AI projects completed.

Learn more about multi‑agent systems.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.