Organizers publishes winner write — ups on May 13, 2026 documenting architectures, data pipelines and lessons learned.

A developer — focused hackathon produced operational AI agents built with Agent Builder, Elasticsearch, ES|QL and Elastic Workflows that automate messy workflows, run domain — standard analytics, and reveal real-world errors;

The Elasticsearch Agent Builder Hackathon produced several working AI agent systems that demonstrate how LLM reasoning plus structured tool access can automate complex operational tasks. Organizers published winner write — ups on May 13, 2026 documenting architectures, data pipelines and lessons learned. These artifacts aim to show developers and teams how to move agents from prototypes toward production. Judges rewarded entries that went beyond single — query assistants to perform multi — step reasoning, execute statistical analysis inside ES|QL, and trigger downstream integrations such as Slack, Jira and email. Teams were explicitly encouraged to pair Agent Builder — style LLM reasoning with orchestration features like Elastic Workflows and search to connect disconnected systems and automate internal processes.

PHAROS, developed by Prajwal Sutar, is a four-agent pharmacovigilance pipeline that ingests FDA adverse event reports, performs WHO-standard statistical analysis entirely inside ES|QL, and generates regulatory paperwork in JSON. The system issues alerts to Slack, Jira and email and completes the full chain in under a minute. Sutar’s write — up explains the choice to keep statistical computation inside Elasticsearch and the JSON-parsing steps used to integrate notifications and tracking.

Gauntlet, by Kavish Sathia, is an adversarial testing framework that deploys a mocking agent to intercept a primary agent’s tool calls and try to break its logic. It uses a two-memory architecture so adversarial strategies evolve across repeated runs, and the builder reported unexpected behavior from ES|QL completion functions discovered during testing. Notably, Gauntlet was rebuilt from scratch 48 hours before the deadline, underscoring iterative pivots common in agent development.

Fredrick Kioko’s three — agent duplicate — detection system targeted a practical public — health problem in Kenya’s HIV testing infrastructure. The chain scanned 1,010 anonymized records in under 10 seconds and surfaced 131 duplicates, including same-day, multi — facility cases that would have taken weeks to identify manually. The post stresses that explainability and traceable decision logs are essential for clinical and operational adoption.

Each published winner post includes architecture diagrams, implementation notes and takeaways intended to help other builders replicate or adapt the approaches. Authors and organizers emphasize that any productization or release timing for described features remains at the company’s discretion; the materials are presented as demonstrations of what combining LLM reasoning with search and orchestration can achieve.