AI Coding Product Evolution: From Autocompletion to Human‑Agent Collaboration

In 2024 Kuaishou launched an AI coding tool focused on code completion (tab‑completion). Early experiments showed a generation rate of only 5%–20% despite heavy investment (≈90% of engineering effort). By integrating a simple ChatGPT‑style sidebar query feature, contribution rose dramatically with minimal effort, demonstrating that asking the model for larger code blocks yields higher productivity.

From 2024 to 2025 the product shifted to a Coding Agent that could execute tasks autonomously. Teams reported penetration rates climbing from 70%–80% to near‑full adoption, and in the last four weeks some teams achieved >90% and even 100% code‑generation ratios. However, the serial “ABAB” interaction pattern—spending minutes typing a request, then waiting ten minutes for the Agent to run, then returning to fix errors—remained painful and fragmented developer flow.

To address this, the team introduced a spec‑driven long‑running Agent loop . By externalizing specifications, progress, execution history, and constraints (the “specs”) outside the conversational context, the Agent can run continuously without repeated human prompts. The Harness environment enforces machine resources, permissions, and project constraints, allowing the Agent to act like a long‑running service. As model capabilities grew, many engineering tricks (e.g., Fast Apply, vector‑based code‑base indexing) became unnecessary; simple diff‑based replace operations now achieve >90% accuracy without heavyweight services.

State externalization also enables treating any compute target—container, cloud instance, or local machine—as a uniform execution node . Installing an Agent runtime and daemon on the node lets it be swapped or scaled transparently, supporting large‑scale Agent Swarm scenarios. This abstraction removes the local/cloud dichotomy and supports seamless migration and failure recovery.

The final step is achieving human‑agent equal collaboration . By exposing the same task list through a CLI, both humans and Agents share context, eliminating information monopolies. Agents can now see the same work items, split tasks, and @‑mention humans when needed. The discussion also highlights broader organizational challenges: test quality degradation, on‑call hand‑off, and the need for shared artifacts (logs, traces, documentation) to keep Agents aligned with human intent.

Overall, the presentation argues that moving from ad‑hoc autocompletion to a spec‑driven, state‑externalized Agent architecture, and finally to a shared work plane, is essential for turning personal productivity gains into measurable organizational efficiency.