OpenAI Codex Failed to Complete a Single Run in Six Days on Open-Source AI Agent

·15 views

A developer gave OpenAI's Codex six days to work on Goose, an open-source local AI agent, but the tool never completed a single full run. Rather than advancing the actual feature, Codex repeatedly over-engineered the project's quality gates, turning a ten-minute build cycle into one lasting over an hour without catching any new defects. The agent also applied unprompted security hardening — including threat modeling and defensive layers — to a single-user local development tool that had no such requirements. Notably, the same underlying model, GPT-5.6 Sol, performed well on the same codebase when run through Goose's own engine instead. The author concluded the failure was not a model problem but a harness problem, attributing the breakdown entirely to how Codex orchestrated the work.

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)

FlowScript 0.1 Debuts as a Semantic Language for Modeling Apps Before Coding

Developer Erland Kjensli has publicly released FlowScript 0.1, an experimental domain-specific language designed to describe the semantic model of an application before any implementation begins. The project addresses limitations of traditional flowcharts, which struggle to capture the full complexity of modern apps including state, navigation, UI structure, and dataflow. FlowScript separates structural hierarchy, navigation graphs, and state graphs as distinct but related concepts, aiming to reduce ambiguity in application modeling. A key hypothesis behind the language is that responsive UI layouts can be treated as projections of a single underlying semantic model rather than separate application hierarchies. The initial release is available as a working draft on GitHub under version 0.1.0.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Encoding API contracts in tool schemas beats prose descriptions for smaller LLMs

A developer team discovered in May 2026 that two AI models — Opus 4.7 and Sonnet 4.6 — produced different JSON outputs from the same tool description, with the smaller model returning an incorrect structure that triggered downstream 422 errors. The root cause was not a capability difference but an exposure gap: the larger model had seen the proprietary field names during training, while the smaller one guessed plausible-sounding alternatives. Because tool descriptions are prose rather than enforced contracts, models without prior exposure to a custom schema have no validated structure to follow. The team resolved this by migrating the structural contract into the JSON schema itself, using discriminated oneOf branches with strict required fields for the nine most common pattern types. They applied a permissive fallback branch for rare pattern kinds, limiting the upfront effort to roughly one afternoon while targeting the failures that actually occurred in production.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Discovers AI Agent Was the Cause of Missing Disk Space

A developer recently shared an ironic experience where an AI agent was tasked with investigating missing disk space on their system. The investigation revealed that the agent itself was the culprit, having consumed the storage it was sent to find. The anecdote quickly gained attention in developer communities for its humorous and self-referential nature. It highlights a growing concern about the unintended resource consumption of autonomous AI agents. The incident has been dubbed a distinctly modern tech problem, reflective of how AI tooling is evolving in 2025 and beyond.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Oracle AI Agent Studio's Retirement Assistant Simplifies Fixed Asset Retirement Process

Oracle AI Agent Studio offers a Retirement Assistant that replaces the traditional multi-step fixed asset retirement workflow in Oracle Fusion Cloud with a conversational interface. The tool allows accountants to retire assets by typing free text or uploading a CSV file, eliminating the need to navigate multiple menus or work areas. The assistant validates inputs in real time, identifying missing fields such as retirement dates and flagging errors with specific, actionable guidance before any changes are posted. It works alongside a companion agent called the Retirement Request Assistant, which lets business users submit retirement requests that are then routed to accountants for processing. Together, the two agents form a front-office and back-office pairing designed to streamline the end-to-end asset retirement workflow.

0 comments Read more at DEV Community