Frontend · Backend · Artificial Intelligence

Harness + SDD + Multi-Agent: A Production-Proven Full-Stack AI Development Workflow

By 得物技术 · May 7, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Why it matters

This isn't another prompt engineering guide — it's a production-tested system for making AI-generated code actually mergeable. The Harness + SDD + multi-agent pattern addresses the real bottleneck in AI-assisted development: not code generation speed, but code quality and review friction. For teams scaling AI coding, the implicit features warning alone is worth the read.

Summary

Dewu Technology's engineering team has published a detailed methodology for full-stack AI development that moves beyond simple prompt engineering. The core insight is "Harness thinking" — instead of asking AI to create code from scratch, developers give it an existing implementation to imitate, dramatically improving code style consistency and review adoption rates.

The workflow combines three elements: Harness constraints that anchor AI generation to existing code patterns, SDD (Specification-Driven Development) documents that formalize frontend and backend designs before code generation, and multi-agent parallelism that lets frontend and backend agents work simultaneously. The team reports a 50%+ efficiency gain on a typical admin panel feature, compressing a 2+4 person-day effort into 3 person-days.

The article also surfaces a critical pitfall: AI-generated code often contains "implicit features" — behaviors copied from reference code that aren't documented in SDD specs. The team recommends treating SDD as a starting point, not a contract, and reviewing generated code for these hidden behaviors.

Key takeaways

— Harness thinking means giving AI an existing implementation to imitate, not asking it to create from scratch — this dramatically improves code style consistency and review adoption rates.

— Placing frontend and backend code in the same Cursor workspace enables cross-repo Codebase Indexing, letting AI understand relationships between API calls and return structures.

— Full-stack SDD requires two separate documents (frontend and backend) with explicit interface contract alignment and field mapping.

— A proven prompt template for full-stack SDD generation includes workspace context, reference entry points, and a requirements checklist for both sides.

— Multi-agent parallelism (Cursor tabs or Claude Code Subagents) lets frontend and backend code generation run simultaneously, cutting development time.

— A three-phase verification strategy — frontend mock, backend compile, then integration — isolates issues early and reduces debugging overhead.

— AI-generated code often contains undocumented 'implicit features' copied from reference code that developers must actively review for.

— The methodology compressed a 2+4 person-day feature into 3 person-days, a 50%+ efficiency gain.

— Backend code can be validated with just 'mvn clean compile' — no need to start the full service locally.

— Mock data must match backend SDD field types exactly and cover edge cases (empty lists, nulls, long strings) for effective frontend self-testing.

Our take

The real bottleneck in AI-assisted development isn't code generation speed — it's the gap between 'code that runs' and 'code that merges.' Harness thinking directly addresses this by anchoring AI output to existing team conventions.

The implicit features problem is a double-edged sword: AI's ability to replicate undocumented behaviors from reference code can either save time (if the behavior is desired) or introduce subtle bugs (if it isn't). This makes code review more, not less, important in AI workflows.

The 50% efficiency gain is impressive, but the article's honest accounting of 'environment setup, troubleshooting, integration, and self-testing time' suggests the real savings come from parallelization and reduced rework, not raw generation speed.

The choice between Cursor and Claude Code isn't about which is 'better' — it's about workflow preference. Cursor's multi-tab parallelism suits IDE-centric developers, while Claude Code's Subagent model fits command-line and CI/CD workflows.

The SDD approach implicitly acknowledges a limitation of current AI coding tools: they need structured, formalized specifications to produce consistent multi-file changes. This mirrors how human teams operate, suggesting AI coding maturity is about process, not just model capability.

Concepts & terms

Harness Thinking

A prompting strategy where AI is given an existing code implementation to imitate, rather than asked to create from scratch. This ensures generated code matches project conventions in naming, structure, and layering.

SDD (Specification-Driven Development)

A workflow where AI first generates formal specification documents (proposal, spec, design, tasks) for both frontend and backend before writing any code. These documents serve as the blueprint for parallel agent execution.

Codebase Indexing

Cursor's feature that performs vector embedding on all code in a workspace, building a semantic index. This allows AI to retrieve relevant files across repositories without manual file specification.

Multi-Agent Parallelism

Running multiple AI coding agents simultaneously — typically one for frontend and one for backend — to generate code in parallel after SDD documents are finalized. Supported by Cursor's multi-tab interface and Claude Code's Subagent mechanism.

Implicit Features

Undocumented behaviors that AI automatically copies from reference code when generating new features. These can include form clearing, data transformations, or default value logic that exists in the reference but isn't specified in the SDD.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗