Developer finds four critical bugs in AI emergency response loop built for hackathon

·1 views

A developer built CrisisMesh, a seven-agent AI fleet on Google ADK and Gemini Flash, designed to coordinate school emergency responses for the All Things Agentic Hackathon. The system exposed four interconnected bugs in its automated follow-up loop, which was meant to ping unresponsive staff and escalate to floor wardens. Among the failures: escalation messages were sent to the very person being searched for, the loop never stopped paging wardens, placeholder Slack IDs inflated reachability counts, and a tick guard ran per-process rather than across distributed instances. The bugs all passed unit tests because each function worked correctly in isolation — the failures only emerged at the integration and system level. The developer resolved the issues by rewriting tests to assert real-world outcomes rather than return values, and by using Firestore leases to coordinate state across multiple containers.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Adding a Fourth AI Model Mid-Test Revealed Flaws a Three-Model Setup Would Have Hidden

A developer running a field test for an AI debate validation tool called AdversarialDebate added a fourth language model halfway through the experiment after realizing the original three-model setup could not adequately test the full diversity spectrum. The initial lineup of GPT-4o-mini, Gemini 2.5 Flash, and DeepSeek-V3 only covered pairings between US and Chinese labs, leaving no cross-continental comparison at the far end of diversity. Adding Mistral Small 3.2 introduced a China-EU pairing that became the highest-scoring combination, with a 0.982 average score and 97% verdict rate. The expanded setup also uncovered that weak model diversity could perform worse than no diversity at all, and that the top-scoring pair was not necessarily the safest due to capitulation cascades. The author concluded that a field test's primary value lies not in generating numbers but in determining whether an experiment can actually answer its core question.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Micro-Frontend Template Series Concludes With Ownership Model and Fork Guide

The final installment of a five-part developer series on a micro-frontend platform addresses team ownership and how organizations can adopt the template for their own use. Each platform component — Host, Store, domain MFEs, and the manifest registry — is assigned to a specific team, with domain teams restricted from importing each other's code to maintain a clean two-level dependency tree. All organization-specific settings such as branding, identity provider configuration, and manifest URL are centralized in a single platform.config.json file. The platform is distributed as a GitHub template repository, designed so adopters can fork it, edit that one config file, and point it at their own infrastructure without modifying core logic. This approach is intended to keep forks upgradable from the upstream template over time, with versioned contracts serving as stable seams for pulling in future changes.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why One Tech Team Chose Webpack Module Federation for Enterprise Microfrontends

A software engineering team has published a multi-part series detailing the architecture of a real enterprise microfrontend platform built using Webpack 5 Module Federation. The core requirement driving the design was enabling any team to deploy changes independently, without triggering a rebuild or redeployment of the central Host application. To achieve this, the Host loads a runtime manifest file that maps page routes to remote bundle URLs, meaning adding or updating a team's page requires only a manifest entry change. Shared dependencies such as React and the design system are declared as singletons, ensuring all microfrontends use the same running instance rather than separate copies. The first installment focuses on why Module Federation was selected over two other viable alternatives, with subsequent parts covering auth, CI/CD, and safe manifest updates.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Adam and AdamW Became the Backbone of Modern LLM Training

Adam, the optimizer now central to training large language models, was originally proposed by Diederik Kingma and Jimmy Ba in December 2014 — well before the Transformer architecture existed. It works by maintaining adaptive per-parameter learning rates using exponential moving averages of gradients, allowing billions of parameters to update at different effective rates without exploding or stalling. The Transformer paper adopted Adam directly in its training recipe, cementing its dominance in deep learning. However, a subtle flaw in how Adam handles weight decay led Ilya Loshchilov and Frank Hutter to propose AdamW in 2017, which decouples regularization from gradient statistics for more consistent weight shrinkage. AdamW is now the standard choice for Transformer training, and Adam's decade-long influence was formally recognised with an ICLR Test of Time Award in 2025.

0 comments Read more at DEV Community