--- title: "From a shopping app to open source: JD ships an agent and a coder" date: 2026-11-09 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-jd-joyagent-coding language: en --- # From a shopping app to open source: JD ships an agent and a coder > At its 2025 tech conference JD.com open-sourced JoyAgent, a production-grade AI agent platform, alongside JoyCode, an AI coding assistant, both free to deploy. A company best known for delivering boxes to your door decided it had something else worth giving away: the software that runs its own internal AI workers. JD.com, one of China's largest retailers, walked onto the open-source stage in July 2025 with two tools it says it already uses across the business. The bet is that if outside developers build on JD's agent and coding platforms, the whole ecosystem gets cheaper to grow — and JD stays at the center of it. ## Key takeaways - JD Cloud open-sourced JoyAgent, described as a 100-percent product-level open agent platform covering frontend, backend, framework, and core sub-agents. - Alongside it came JoyCode (JoyCode 2.0 / JoyCode-Agent), an enterprise coding assistant built for large, messy codebases. - Both were unveiled at JD's 2025 global tech conference and positioned as the "agent + coding" twin engine for serious development. - The moves extend JD's JoyAI brand, upgraded at WAIC 2025, which spans models from 3B to 750B parameters across language, voice, image, and video. ## Why a retailer is shipping AI infrastructure JD moves more than goods. It runs warehouses, logistics, a health arm, and a sprawling supplier network, all of which throw off problems that fit AI agents well: pull data from five systems, decide something, act on it. The company says it has already built tens of thousands of stable production-grade agents inside its own operations. Rather than keep that machinery private, JD packaged the platform and opened the source. The stated pitch: smaller companies get a battle-tested agent system without building one from zero, and JD gets a wider developer base feeding improvements back. ## JoyAgent: an agent that is not a toy Most "agent" demos fall apart the moment they touch a real company's tangle of databases and approvals. JD's answer, JoyAgent 3.0, is built around multi-agent coordination — several specialized agents handing work to each other — plus a dynamic execution engine for tasks that need ordering and branching. It also includes data tooling so the agent can actually query a company's tables and documents. JD says JoyAgent scored in the global first tier on GAIA, a benchmark that tests whether an agent can finish messy real-world tasks using tools. The company reports validation and test accuracy above 77 percent and 67 percent respectively. Those are JD's own reported numbers, and GAIA rankings shift as teams submit, so treat them as a snapshot, not a permanent medal. Since the July 2025 open-source release, JD says the project passed 10,000 GitHub stars, a rough signal that developers at least showed up to look. ## JoyCode: the agent that writes and fixes code The second half of the pair is JoyCode, an assistant aimed at the unglamorous reality of enterprise software: million-line codebases nobody fully understands. JD says its JoyCode-Agent reached a 74.6 percent pass rate on SWE-Bench Verified, a respected benchmark where models must actually resolve real GitHub issues. The company frames that as a top-three global result, again a self-reported claim worth checking against the live leaderboard. The deeper idea is symbiosis. JoyAgent can call JoyCode's capabilities; code JoyCode generates can flow back to improve the agent. JD describes a loop where agents and coding logic evolve together instead of living in separate products. ## Where this touches ordinary life JD's AI is not abstract. Its digital-human avatars, built on the same JoyAI foundation, already pitch products in livestreams for thousands of brands. During one recent shopping festival the company said its digital humans lifted a partner's livestream sales and cut operating costs sharply — numbers that come from JD, so read them as promotional. The point stands regardless: agents that started inside a retailer are now leaking into storefronts and call centers everywhere. This is also where embodied intelligence (具身智能) enters the story. JD demonstrated humanoid (人形机器人) and service robots running on its models, a reminder that a "model" column item often becomes a body in a warehouse. ## Honest limitations Nearly every headline number here — GAIA scores, SWE-Bench rank, star counts, sales lifts — originates from JD's own announcements and conference talks, not independent testing. Open-source code is real and inspectable, but "open" does not mean "proven best in the world." Running JoyAgent or JoyCode at scale still demands real engineering and compute, and the 750B flagship model is a commercial product, not part of what was given away for free. Benchmark wins on curated tasks do not guarantee smooth behavior in a company's unique mess. ## What readers can do now - If you run a small team, clone JoyAgent from GitHub and prototype one internal workflow before buying a closed agent platform. - If you are a developer, try JoyCode-Agent on a real bug in your repository and compare it honestly against the tools you already pay for. - If you are a shopper, notice when a "person" on a brand livestream is a model-driven avatar, and decide how much that changes your trust. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-jd-joyagent-coding Free to quote with attribution and a link to the original.