Developer Debugs AI Agent Loop, Adds Run History After Silent Script Failures

·1 views

A software developer building an AI-powered ticket-execution agent encountered a silent failure on day three when a shell script could not locate user stories due to a filename prefix mismatch between 'US-04' and the actual 'NN_slug.md' file naming convention. A secondary bug caused an overly strict pipefail setting to kill the script before its own error message could display, and a content-matching fix later revealed a third issue where grep matched ticket IDs inside unrelated file bodies rather than headings. To address the lack of run visibility, the developer added timestamped JSON log files for each agent run using a tee pipeline, ensuring no output would be lost between sessions. Over the following day and night, 29 runs were logged across 20 user stories, accumulating nearly 500 turns and approximately USD 55 in API costs, with one ticket requiring four attempts before completing successfully. A final status-detection bug — where the script matched the word 'BLOCKED' anywhere in output rather than only in the status line — was caught and fixed manually after the new history system made it visible.

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)

Anthropic's 5th-Gen Models Prompt Better With Explicit Contracts and Context

DEV Community has published a practical guide outlining updated prompting best practices for Anthropic's fifth-generation models, including Claude Sonnet 5, Opus 5, and Haiku 5.1. The guide is aimed at cloud engineers and emphasizes providing reasons behind instructions rather than bare rules, and defining explicit output contracts instead of vague requests. Other key recommendations include capping task complexity to prevent over-engineering, placing reference documents before questions in long-context prompts, and using examples to demonstrate style rather than describing it in prose. The guide also advises against including verification scaffolding in prompts and suggests softening overly directive tool-use language inherited from earlier prompt patterns. A full reference document is available via the original DEV Community post.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Guide: Building a Gemini AI Assistant for Android XR Smart Glasses in Kotlin

A technical tutorial published on DEV Community outlines how to build a production-ready AI assistant for Android XR smart glasses using Kotlin and Google's Gemini backend. The architecture separates device-specific XR and wearable APIs from core business logic, using a layered folder structure to isolate SDK changes. The guide covers voice input processing, AI response generation, and audio or projected XR output, with structured concurrency via Kotlin coroutines to avoid blocking the main thread. It also addresses key concerns such as error handling, security best practices, and performance metrics including network and AI latency. Developers are advised to follow current Android XR and Jetpack XR documentation closely, as the relevant APIs are still evolving.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a 9-Agent AI Pipeline Delivers Working Code in Under 4 Minutes at Scale

A development team built a nine-agent AI pipeline capable of converting a natural language use case into working code, an interactive preview, and implementation documentation in under four minutes. The system serves 800 to 1,000 users daily, with each request originally consuming around 30,000 tokens and triggering 15 to 20 model calls. Engineers divided the workflow into specialized agents — covering analysis, code generation, evaluation, refinement, and documentation — each assigned a narrow role with defined inputs and outputs to isolate failures and improve reliability. Two stages, a Schema Validator and a Documentation Builder, were implemented as deterministic programs rather than AI calls, eliminating two potential hallucination sources and reducing token use to zero for those steps. The core lesson drawn is that production-grade multi-agent systems depend not on adding more agents, but on clearly distinguishing which tasks require AI reasoning versus which are better handled by conventional software logic.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

HTB Tactics Walkthrough: Windows SMB Box Compromised via Blank Admin Password

A Hack The Box (HTB) challenge named 'Tactics' presented a Windows machine accessible over a network. Nmap scanning revealed open ports 135, 139, and 445, confirming the target was running the SMB protocol. The attacker used the Impacket tool 'impacket-psexec' to attempt remote shell access over SMB. Surprisingly, the Administrator account was found to have no password set, allowing immediate authentication without any credential cracking. The challenge flag was then retrieved directly from the Administrator's Desktop directory.

0 comments Read more at DEV Community