31 Teams Push AI Agents Forward in 48‑Hour Beijing Hackathon

Why the Hackathon?

With AI tools making one‑sentence app generation easy, the core question became how to evolve AI agents from personal toys into enterprise‑grade tools. The "Beyond Prompt: Agents in Action" hackathon held at Microsoft Asia R&D headquarters gathered 31 teams and nearly a hundred developers to test whether agents could truly enter business workflows and solve real problems.

Key Challenges Identified

Organizers highlighted three pain points: agents tend to work in isolation, lack domain expertise, and struggle to find genuine problems to solve. The event’s goal was to push developers beyond prompt‑writing into tackling enterprise‑level integration.

Three Competition Tracks

Tracks were chosen to reflect the most urgent enterprise use cases for agents:

Coding Agent – addressing collaboration, code review, and mobile development gaps.

Vertical Data‑Analysis Agent – handling unstructured data in e‑commerce, finance, legal, where generic LLMs fall short.

To‑B Open Scenarios – from HR to supply‑chain, defining real problems for AI to automate.

Teams quickly selected directions, ranging from large‑tech engineers to academic researchers, all aiming to deliver functional agents within 48 hours.

Technical Foundations Provided

Before coding began, Xiaoshu Technology presented its Agent API stack, which includes persistent memory, a secure code sandbox, and end‑to‑end task delivery, as well as an AI‑powered search service covering 30+ languages and 4.8 billion scholarly documents. These services were positioned as the "water, electricity, and gas" needed for agents to operate.

Highlights from Winning Projects

First Prize – TeamLoop : Added a control plane on top of Codex and Claude Code to enable task assignment, evidence collection, and approval, turning a solo coding agent into a collaborative, auditable digital employee.

Second Prize – Token Matrix : Built a data‑center operations agent that maps device knowledge, SOPs, and historical faults into a knowledge graph, automatically recommends solutions, creates tickets, and validates closures.

Second Prize – Rainbow Wireless : Automated bid‑tender scouting by aggregating sources, understanding content, evaluating relevance, prioritizing leads, and emailing results.

Third Prize – Dataworks : Delivered a one‑stop data‑analysis platform where non‑technical users can issue natural‑language queries to extract, analyze, and visualize enterprise data.

Third Prize – Research Screening Agent : Reduced clinical trial patient screening time by 90% and increased throughput tenfold using multimodal document understanding.

Third Prize – Beyond CoAgent : Provided a collaborative console for multiple agents to coordinate and manage operations tasks.

Additional awards recognized commercial value, technical innovation, and community impact, such as an AI‑driven bidding decision platform, a digital employee replica with hand‑off to humans, and a training‑operation agent for coaches.

Insights from Workshops and Panels

Guest speakers emphasized that hiring now values demonstrable AI‑agent solutions over mere tool familiarity, warning of "penetration‑rate traps" and "scenario extinction risks." They advocated a design principle: let AI handle large‑scale information retrieval and repetitive execution, while humans focus on deep communication and critical judgment.

Product leaders from Xiaoshu stressed that without real‑time web search, agents become "digital fossils," and that their Agent API’s thread‑aware execution engine, tool‑dispatch center, secure sandbox, and streaming output enable a single call to cover the full information‑to‑delivery pipeline.

Participant Reflections

Across the board, teams discovered that the hardest part was not improving model intelligence but defining clear boundaries for agents. Successful projects limited agents to narrow, well‑defined tasks—such as information retrieval and structuring—rather than attempting end‑to‑end solutions.

Examples include TeamLoop’s approval layer, Token Matrix’s diagnostic‑first approach, and Replica’s explicit human hand‑off for unanswered queries.

Overall Takeaways

The hackathon demonstrated that real‑world AI agents must evolve from answering questions to completing tasks, starting with a disciplined scope. Agents add value by offloading large‑scale data processing, freeing humans for higher‑order work, and delivering measurable economic impact.

Future Outlook

Organizers view the Beijing event as the first step in a series of "Beyond Prompt" gatherings, aiming to continue refining AI‑agent infrastructure, scaling secure, efficient services, and expanding enterprise adoption.