Taobao Overseas Tech's AI-Native R&D Shift: SDD, Agents & 85% AI Coding

AI Coding Evolution and SDD-Driven Initial Results

AI coding has evolved from line-level IDE completions to conversational assistants and now to Agent-driven autonomous coding. The team adopted Specification-Driven Development (SDD): write design specs in natural language, then let AI generate code from specs and the existing codebase. Specs become the "instruction"; code is the artifact, not the starting point.

Quality is ensured through a phased gating process with human reviews at six stages: ① Project Preparation → ② Requirement Clarification → ③ Technical Proposal Generation → ④ Proposal Execution → ⑤ Test Execution → ⑥ Closure & Archiving. AI coding share rose from 5% to 85%, and overall delivery efficiency continues to improve.

The Perception Gap: Individual Speed vs. Organizational Throughput

Despite 85% AI-generated code for both frontend and backend, business and product stakeholders still feel delivery hasn't accelerated. Data shows coding occupies only ~20% of the end-to-end lifecycle; a 200% coding speedup yields only ~10% overall gain. Most time is spent in requirement alignment, PRD writing/review, technical design/review, integration testing, and cross-team coordination. The real bottleneck lies in the serial "Business → Product → Tech" handoff chain.

Agent-Driven Paradigm: Unified Interface, AI Execution, Human Decision

Core logic: align business/product/tech work surfaces, reduce communication overhead. DingTalk (ChatBox) is the proven collaboration paradigm; embedding an Agent brain turns it into an end-to-end execution platform. The traditional serial flow (Business requests → Product designs → Tech delivers) is reconstructed into "AI executes + Human decides" within a single DingTalk group.

Overall Solution

One-sentence trigger, avatar auto-relay: Operations states a business direction in natural language in the group; four AI avatars (Product, Tech PM, R&D, Analyst) automatically pick up and run the full chain from clarification to coding to data analysis.

Humans only at key gates: Three human checkpoints — PRD confirmation, AI test acceptance, final human sign-off. Avatars "get things done"; humans "judge correctness and releasability".

Clear engineering path: Every stage has a defined Agent carrier, dedicated knowledge base, and required engineering capabilities — mature components, not a demo.

Data feedback loop: Post-launch business data flows back; Analyst avatar generates strategy suggestions, triggering the next iteration.

Engineering guardrails throughout: Collaboration bus, artifact traceability, cross-review, mandatory gates for high-risk actions ensure controllability, trust, safety, evolvability.

Core Process

DingTalk group as hub; AI executes, humans decide. Agent drives requirement → design → clarification → delivery in a closed loop. All documents, meeting notes, and decisions stay in the group, forming a living domain knowledge base that continuously improves delivery speed and quality.

Three key implementation paths: ① Unified work interface ② AI-driven process collaboration ③ Automated delivery verification.

Key Paths

Channel: DingTalk group is the single work surface and context carrier; all docs, emails, meetings settle directly in the group.

Roles: AI produces (drafts, retrieves, summarizes, codes); humans decide (clarify, trade-off, approve, canary).

Flow: Requirement → Design → Clarification/Review GATE → Delivery → Release GATE — all executed inside the project group.

Targets

Focused on end-to-end delivery efficiency, with result and process metrics.

Result metrics (by complexity):

Faster: Small needs (≤5 person-days) zero R&D investment, Agent end-to-end same-day release; medium/large needs 20-30% cycle compression vs. baseline.

More: Same headcount, 40% monthly delivery throughput increase, validating capacity release from zero-R&D small needs.

Quality floor: Bugs per KLOC no worse than current; online incident and rollback rates no worse than baseline.

Process metrics (paradigm adoption health):

Paradigm penetration: % of needs delivered via new paradigm; Phase 1 target 100% for small needs in merchant supply domain, 50% for complex, then expand.

AI draft adoption: MRD/PRD final version evolved from AI draft (not rewritten), target ≥80%.

Clarification/review efficiency: Auto-instrumented group telemetry (rounds, gate durations) to baseline and reduce; MRD/PRD currently manual, will measure after AI assist.

AI coding share: Small needs from ~90% to 99-100% (near full automation); medium/large from 60% to 90%.

Key Capabilities

2.2.1 Unified Work Interface

DingTalk group integrates plugins, knowledge bases, databases. One project = one group; created at kickoff. Business, PD, PM, R&D, QA, and AI Agents all join. Group messages, files, cards auto-form project memory. Agents participate as group bots on equal footing.

2.2.2 AI-Driven Process Collaboration

AI drives, humans decide; process pulled back into the group. Business, PD, R&D collaborate in parallel. AI Agents clarify requirements, generate MRD/PRD drafts and technical designs, auto-archive discussions and decisions. Rhythm shifts from low-frequency big reviews (meeting → wait for doc → meeting) to high-frequency micro-cycles (AI draft → human quick review → AI immediate iterate). For complex coding, developers still prefer local IDE; the process guides them back to IDE for vibe coding, then return to group for updates.

Two tracks by complexity:

≤5 person-days: coding, testing, acceptance fully in-group; AI produces Diff and Test Report, aiming for one-shot success.

>5 person-days: group focuses on MRD/PRD/design review, upstream/downstream sync, progress tracking; vibe coding done locally in IDE, then results posted back.

Unified principle: regardless of track, requirement progression, artifacts, and key decisions stay in-group, flowing as card messages with instant traceability.

2.2.3 Automated Delivery Verification

Under SDD, PRD and technical design merge into machine-executable Spec. AI runs coding, unit tests, regression, deployment, verification, release automatically. Humans guard only three gates: design approval, core code review, canary release approval.

Two tracks again:

≤5 person-days: full closed-loop in DingTalk, no context switch, no queue.

>5 person-days: DingTalk as process cockpit; vibe coding and deep testing/regression back in local IDE — keeps automation speed while retaining control for complex engineering.

Every pipeline stage hangs on objective quality gates — unit test, lint, security, regression, online metrics — anomalies auto-rollback. Human gates are risk-graded from L1 (lightweight) to L4 (full human decision). AI only computes impact scope; decision authority returns to humans.

Practice Path and Results

2.3.1 Practice Path

Pilot scenario selection: Start with "what-you-see-is-what-you-get" simple needs — copy/button tweaks, sorting rule changes, config updates. Simple PRD, enumerable design, low acceptance bar; easiest to validate "one sentence → release" in-group loop. Then expand to medium-complexity features.

In-group gate mechanisms: Standardize "permission check, risk identification, output quality" as AI gates in-group: every Yes/No must verify approver identity; AI self-checks drafts (completeness, conflicts, missing deps) before human review; high-risk changes force human decision.

Measurement & feedback loop: Every group flow instrumented — stage latency, AI draft adoption, human rework count, final quality. Data feeds back into Skill prompts, process templates, forming "real business runs data → data improves AI → capability expands pilot scope" virtuous cycle; also seeds digital avatar datasets.

2.3.2 Phase Results

New paradigm validated in Supply, Payment, and Marketing domains. Capability extended from code assistance to requirement analysis, design generation, R&D execution, project coordination — proving end-to-end delivery and cross-role collaboration feasibility.

R&D Delivery: Supply domain small needs achieved full in-group delivery from MRD/PRD/design through coding, deployment, test, release. AI corrected technical design via code analysis, cutting overall R&D estimation by 50%. Complex needs (e.g., Panama) use hybrid "group collaboration + IDE development" mode, supporting multi-module CR, test case generation, pipeline tracking, pre-prod issue triage.

Project Coordination: Payment and Supply domains connected meeting minutes, PRD clarification, todo tracking, technical design, R&D delivery. Meeting conclusions auto-convert to action items with owners and deadlines, flowing into design, CR, code delivery — continuous sync of progress, risks, todos, reducing scattered info and repeated alignment.

Requirement Analysis: Marketing domain completed current-state mapping, feasibility judgment, upstream/downstream impact analysis on real needs, generating PRD drafts and interaction demos. Product, business, R&D discuss on same analysis artifact, cutting early-stage back-and-forth.

Future Iterations

Process End-state: Operations initiates in group; multi-role AI avatars auto-relay through clarification, design, dev, release, data analysis; humans only at final acceptance; business enters continuous self-iterating loop.

Data Flywheel: Post-launch data auto-feeds back, calibrates knowledge bases, optimizes avatar behavior, evolving system from "one-off delivery" to "continuous self-iteration and refinement".

Management & Intervention: Build unified admin console for monitoring, rule definition, strategy analysis, manual intervention — ensuring "observable, controllable, correctable".

Digital Employees: Beyond per-need "avatars", create "digital employees" that reside permanently in a business domain with defined responsibilities and authority boundaries. They proactively claim eligible needs, complete within boundaries, escalate only when out of scope. Avatars own a single delivery; digital employees own continuous delivery for a domain.

Summary: L1–L3 Maturity Levels

AI-era R&D paradigm classified by AI participation into L1–L3. After one month of real delivery practice, Taobao Overseas Tech achieved:

Full-cycle delivery validation; business/product/tech aligned on collaboration flow.

Communication and collaboration efficiency uplift across roles; eliminated fragmentation-induced information gaps.

R&D paradigm upgraded: ≤5 person-day needs fully delivered in-group (coding, deploy, test).

Small needs essentially at L3; most needs still at L2. Next focus: extend L3 to medium/large needs via four workstreams:

Process Specification: SDD spec as input, fused with AAIC Skill hosting/scheduling to drive "coding & integration & test", gradually automating parts that currently require local IDE; humans keep only risk-graded gates.

Document Quality: Iterate prompts, example libraries, project/domain knowledge bases so MRD/PRD/design AI drafts reach review-ready quality, reducing rewrites.

Engineering Foundation: Polish Skill orchestration, bot latency, group card interactions; complete identity verification for every Yes/No and cross-org authorization.

Asset Iteration: Telemetry on stage latency, AI draft adoption, human rework; group info沉淀 as digital assets — traceable, replayable, reusable — continuously enriching and calibrating knowledge bases.

When these four flywheels spin, the L2/L3 boundary truly sits on scheduling authority: process engine owns stage transitions; AI autonomy expands from single-stage to cross-stage end-to-end; humans arbitrate only when risk materializes. Metrics shift from AI code share to end-to-end cycle time — throughput no longer linearly tied to headcount, but to knowledge base quality, domain Skill coverage, and trustworthiness of permission/boundary controls.