One Engineer, Three Platforms: AI-Native Mobile Development Without Cross-Platform Frameworks

Background: The Multi-Platform Native Development Dilemma

Mobile native development has long faced the challenge of delivering features across iOS, Android, and HarmonyOS. Three pain points dominate:

Architecture side: Cross-platform frameworks (React Native, Weex, KMP, Rust) enable code reuse but introduce framework adaptation, interoperability, and dependency maintenance costs. For example, a reusable camera component must adapt to native permissions and is subject to upstream version changes.

Requirement side: Parallel development for three platforms suffers from resource scheduling and context-switching overhead.

R&D side: Maintaining separate iOS/Android/Harmony teams incurs high coordination costs for design sync and joint debugging.

The article frames this as an "impossible triangle" of flexibility, power, and maintainability — a classic software engineering trade-off. Even at Alibaba, with highly centralized infrastructure, multiple cross-platform solutions coexisted, increasing architectural complexity.

AI-Driven Paradigm Shift: "One Person, Three Platforms"

The emergence of large language models enables a new approach: AI directly participates in native development without a cross-platform adaptation layer. From a single requirement source, AI can perform requirement analysis, solution design, coding, and debugging in parallel for all three platforms. This eliminates the need for a large cross-platform framework, three separate teams, and staggered multi-platform sprints.

As model capability improves and cost drops, mainstream cross-platform frameworks (React Native, Weex, etc.) face disruption. Native multi-platform development becomes a viable alternative; business developers no longer default to cross-platform frameworks.

Evolution of the AI-Native Workflow

From Vibe Coding to Structured Workflow

Initial Vibe Coding handled isolated tasks (code generation, tweaks, debugging) but left end-to-end delivery to manual stitching. To support complete requirements, the team codified platform-specific engineering experience into Skills , exposed them via MCP tools, and combined them with Specification-Driven Development (SDD) into a Workflow covering requirement understanding, code analysis, design, task breakdown, coding, and verification. The key insight: transform individual-AI collaboration into a team-reusable development method.

Harness and Native Flow 1.0

Workflows still suffered from context accumulation, drift, and recovery issues in long tasks. The team introduced Harness to orchestrate tasks via control flow, manage context in phases, and persist task state and verification results — reducing reliance on continuous chat and manual handovers.

Native Flow 1.0 connected requirement analysis, environment setup, design, per-platform coding, and build verification, enabling a single Owner to drive a full requirement. A real case — Alipay Wealth Tab large-screen adaptation — demonstrated feasibility: the three platforms had divergent internal logic and dependencies; the Owner used the model to analyze differences, formulate a unified plan, and iteratively fix platform-specific component and system-API compatibility issues through the workflow.

Key takeaway: The Owner's critical role is identifying which differences affect business outcomes and owning trade-off decisions; AI handles code analysis, implementation, and troubleshooting.

Phase Validation Results (May–Aug 2026)

Progressed from simple to complex requirements, proving viability.

Over multiple real demands, requirement count multiplied while manpower investment dropped 46% .

Local validation confirmed feasibility and efficiency gains, prompting broader rollout.

Scaling at Ant Financial: Harness 2.0

Scaling surfaced three problems: rigid workflows unfit for diverse businesses, scattered assets raising context-assembly cost, and automated testing becoming the new bottleneck after coding acceleration. The team built a unified solution — Harness 2.0 — around three pillars.

Harness Flow: Capability Reuse and On-Demand Composition

Native Flow 1.0 tightly coupled capabilities with requirement state, execution order, and prior artifacts. Running a single Analyse step required entering the full flow; adding a capability meant modifying both orchestration and state management. This limited independent reuse and extension.

Harness Flow decouples capabilities from fixed processes. Teams select and combine development stages, inject domain-specific skills, and share tools and experience. Principle: common capabilities stay stable; business processes retain flexibility. Coupling depth directly impacts onboarding and evolution cost.

Common Bundle: Unified Asset Architecture

Three-platform code, business knowledge, and iteration records were scattered, forcing Agents to repeatedly hunt and stitch context. Common Bundle organizes all three in a single repository:

Three-platform code: Directories ios/, android/, ohos/ for correlated analysis; each platform still builds and debugs independently.

Shared knowledge: Centralized business knowledge, architecture, APIs, and coding standards declared via AGENTS.md / CLAUDE.md with task-based loading.

Iteration records: plans/ for execution state, archives/ for artifacts and historical decisions, reducing repeated clarification and context reconstruction.

Value: When the Agent analyzes code, it simultaneously accesses business rationale and past decisions. Co-location eliminates cross-repo searches and manual copying. Business consistency is further enforced via rules and automated cross-platform checks.

End-to-End Automated Testing: Closing the Pre-Submission Verification Loop

One Owner owning three implementations means more self-test burden. If build-install, device interaction, and regression still require heavy manual work, coding speed gains don't translate to delivery speed.

The team integrated two tools into the workflow:

Native-E2E: Plans test cases and orchestrates verification. Combines requirements, design, and QA-provided cases to define steps, assertions, and platform coverage (happy path, exceptions, boundaries). Executes and generates reports.

Native-CLI: Unifies device connection and operation across iOS, Android, Harmony — install/launch, UI interaction, screenshot/recording, log collection. Supports local devices and cloud device farms, cutting device-prep and model-coverage costs.

Together they chain case planning, execution, fix, and re-verification while preserving evidence. The Owner reviews pass/fail/uncovered items to confirm readiness for QA. Verification velocity must match coding velocity and provide a basis for release decisions.

Results at Ant Financial (Wealth & Insurance)

People development: >80% of terminal engineers became multi-platform experts, simultaneously iterating iOS, Android, Harmony.

Demand coverage: >80% of new business demands use the one-person-three-platforms mode.

R&D cost: Total demand effort reduced by 40%+ vs. previous approach.

Joint debugging: Shifted from three-platform separate debugging to "single-platform debug, others reuse," drastically improving efficiency.

AI code adoption: Native AI Coding acceptance rate >80%, code throughput continuously rising.

Mindset Shift: Toward Full-Stack Agent Engineers

The practice reshaped thinking: reconstruct R&D around efficient AI utilization. Focus extended to the full chain — requirement understanding, design, coding, verification. Core principle: give AI sufficient context, well-defined tasks, and verifiable outcomes; measure success by delivery quality, cycle time, and investment.

Tools and engineering assets must continuously absorb model advances. Workflows organize tasks, assets provide context, automated verification feeds back results. As models evolve, these mechanisms must adapt to avoid becoming new constraints. Tooling should evolve based on real business outcomes.

Engineers transition from single-platform implementers to "Full-Stack Agent Engineers" responsible for business understanding, solution design, Agent collaboration, cross-three-platform delivery, and accountability for key trade-offs and final quality. "One person, three platforms" is a stage-specific manifestation of this capability in native scenarios, demanding continuous accumulation of business judgment, platform expertise, and verification experience.

Technical Deliverables Supporting the Approach

Capability Reuse — Harness Flow

Reuse experience and tools; compose and extend development chains per business need; continuously improve via team co-construction.

Assets & Context — Common Bundle

Co-locate three-platform code, shared knowledge, iteration records; let Agents fetch context per task; reduce repeated search and understanding.

E2E Automated Testing — Native-E2E, Native-CLI

Plan cases, connect devices, execute tests; support fix and re-verification; help Owner verify results and confirm submission status.

These assets enable teams to reuse proven practices and feed new business learnings back into the system, reducing duplicate construction and continuously improving delivery.

Future Outlook

Current practice validates enterprise-grade feasibility in Ant Financial scenarios. To become mainstream globally, "underwater" work remains:

On the surface, this is AI deeply participating in "parallel, multi-platform R&D" under human constraints and guidance. The essence is a shift from "human-intensive intellectual execution" to "human Planning + AI Executing token-intensive collaboration." This will likely become a foundational mode for most R&D domains in the future.