Zero-Based R&D Rebuild: Tencent's AI Pipeline Delivers 68% of Requirements End-to-End

The article references a prior piece by Tencent engineer danteyang arguing that real change lies not in adding AI assistants to each step but in reconstructing the entire R&D workflow. The core evidence comes from a live project at Tencent Security Center running April–June 2026: 254 requirements were processed, 172 (68%) delivered end-to-end by AI, and 51% finished within two hours versus the traditional 3–5 days. Freed capacity shifted to complex strategy and architecture decisions.

Zero-Based Thinking: Question Why the Process Exists

AI excels at execution; humans excel at judgment. Yet much "judgment" is merely compensating for information silos — front-end/back-end separation, interface docs, review meetings, joint debugging. The zero-based question: if AI holds full-chain context and structured specs replace verbal handoffs, can that friction vanish? The answer is not bolting AI onto every stage but rebuilding the process from zero.

Nine-Stage Pipeline: 254 Requirements, 68% End-to-End by AI

The pipeline stages: Requirement Creation → Requirement Review → Feasibility Assessment → Repository Association → Technical Design Review → Automated Development → Code Review → Automated Fix → Test & Deploy → Merge to Main. Each stage is AI-led with quality gates; humans intervene only at critical checkpoints.

Two Quality Gates: Dual Control on Requirements and Technical Design

Requirement Review AI uses interview-style deep dive: restate understanding, then layered questioning to separate "must clarify" from "should clarify", turning vague ideas into executable specs. A PM requirement pool lets a requirement go from rough description to dev-ready in one place.

Feasibility Assessment scores seven dimensions (requirement clarity, repo match, change scope, etc.); ≥60 passes as an upstream guard.

Technical Design AI sorts by dependency graph, slices vertically, controls task granularity, inserts checkpoints. Before key decisions it runs a "nitpick self-review" to surface uncertainties early, ensuring the output is a construction blueprint not an outline.

Surgical Coding and Review Discipline

Three coding disciplines:

Simplicity first — resist AI over-design; over-complex architecture issues dropped ~50%.

Scope discipline — modify only files listed in the design; out-of-scope findings are recorded but not touched ("discover but don't touch").

Test-first on critical path — write tests before implementation; ~17% of requirements exposed logic flaws early via unit test failures.

Code Review AI follows "improvement equals approval": critical issues block merge; suggestions do not. Security review has six hard lines — privilege escalation, auth bypass, SQL injection, etc. — any hit is a critical blocker.

Human role upgrades to: technical trade-off decisions, implicit context injection (architecture background, legacy baggage), anomaly detection and intervention, final value judgment. Review shifts from line-level syntax to intent-level, focusing on cross-module interaction, boundary conditions, and security impact.

Knowledge Flywheel and Capability Boundaries

The platform proactively injects relevant knowledge into AI context per requirement type; injection rate >90%, solving the "unknown unknowns" problem. After each delivery, knowledge auto-condenses: CR issues become pitfall records, technical designs become templates, review processes become PRD templates — 547 assets accumulated, 200+ added in the last 7 days. Three anti-staleness lines: TTL expiry, agent feedback lowering confidence, repo commits triggering architecture refresh.

Capability boundaries form four dynamic quadrants:

Direct do: routine CRUD, bug fixes.

Human-AI collaboration – risk side: DDL needing approval.

Human-AI collaboration – complexity side: multi-file modules.

Not yet supported: fund flows, cross-repo linkage, etc.

Three-Layer Landscape and Organizational Multiplier Effect

Efficiency gains are not at the individual level. Coding faster is additive; real leverage comes from eliminating wait and rework between stages — 40–60% of the delivery cycle. Structured specs replace verbal handoffs, quality gates shift left, process collaboration replaces departmental handoffs, breaking the "people leave, knowledge gone" curse.

Industry three layers:

Bottom 85%: stuck in personal coding red ocean.

Middle 12%: exploring full-process efficiency, cycle compressed 60–72%.

Top 3% (blue ocean): multi-role full-process collaboration — OK platform anchors here.

Future outlook: requirement-to-delivery distance shrinks drastically; coordination roles redefined; team output rises not headcount falls; 2–3 person feature teams become viable. The platform eating its own dog food accelerates this evolution.

When AI handles most routine needs end-to-end, the R&D question shifts from "how to make humans code faster" to "how to use AI to restructure the process and unleash human decision and creativity." Zero-based thinking makes the process invisible; delivery becomes the only goal.