The bar moved past chat
Microsoft's AI Agents Hackathon 2025 is the cleanest signal. Its winners showcase says over 18,000 developers registered and 570 projects were submitted. Projects were judged on innovation, impact, usability, solution quality, and category alignment.
The Best Overall winner, RiskWise, was not a generic assistant. Microsoft described it as a supply-chain risk analysis agent using Python, React/Next.js, Azure AI Agent Service, Semantic Kernel, and SQL to flag risks from shipping schedules, news, and other data for analysts.
Winners orchestrate, they do not just answer
Apollo, the C# category winner in Microsoft's showcase, used a coordinator plus Athena the research engine and Hermes the analyzer. It retrieved web content, stored knowledge in PostgreSQL with pgvector, checked gaps, and synthesized cited reports.
WorkWizee, the Copilots category winner, handled incident-call work by updating Jira, Bitbucket, Confluence, Outlook, and Teams through natural language. That is the difference between an agent demo and a wrapper: the system changes external state.
2026 tracks want tools and MCP
Google Cloud Rapid Agent Hackathon 2026 asks builders to make a functional agent powered by Gemini and Google Cloud Agent Builder that integrates one partner MCP server. The live page lists Arize, Elastic, Fivetran, GitLab, MongoDB, and Dynatrace as partner tracks.
The same page says the agent should reason, plan, and execute tasks under user oversight. It requires a hosted project URL, public open-source repo, about a 3-minute demo video, selected track, and completed Devpost form. That is a clear artifact checklist.
Enterprise agent tracks care about control
UiPath AgentHack is a current 2026 example with $50,000 in cash prizes and a 7-week online build window. Its page says builders must ship working agentic solutions on UiPath Automation Cloud, not a concept or slide deck.
UiPath's tracks reveal what judges want in enterprise agents: Maestro Case for dynamic exception-heavy processes, Maestro BPMN for defined process flows, and Test Cloud for agentic software testing. The page also says solutions that use coding agents through UiPath for Coding Agents get bonus points inside platform usage.
Generic agent contests still need proof
Agentic AI Innovation Challenge 2026 asks for autonomous AI agents that reason, plan, make decisions, and take actions across healthcare, education, productivity, sustainability, accessibility, and open innovation. It asks for project description, demo video, source code repo, screenshots or presentation, agent explanation, and optional live demo.
Its judging criteria are innovation and creativity, technical implementation, real-world impact, user experience and design, and presentation and demo. Those are broad, so the way to stand out is to show a narrow workflow completing end to end.
Build the boring control surfaces
ModelProof, Microsoft's JS/TS category winner, is a good pattern because it supervised AI outputs with dual-model consistency checks and real-time audits for hallucination, bias, toxicity, and intent alignment. The project was about trust, not just generation.
For your own AI agent hackathon project, show logs, state, retries, tool calls, permissions, and handoff points. A judge should be able to answer three questions fast: what did the agent decide, what did it do, and where can a human stop it?
< read by a human · updated as things change >
browse hackathons