Developer finds critical agent bugs by letting hackathon judges write the tests

·1 views

A developer building an autonomous cannabis compliance agent for the All Things Agentic Hackathon discovered a structural flaw common to agent demos: the builder also writes the tests, making success nearly inevitable. To counter this bias, the project stored Kentucky cannabis regulations as structured data and used deterministic code to check compliance, keeping the AI model out of the core safety logic. A fixture service was built allowing judges to generate custom test scenarios the developer had never anticipated, and the first external scenario immediately exposed three bugs in the system. One critical flaw allowed a package with a failing required test to remain movable after escalation, because the agent's escalation logic performed no protective hold. The fix now distinguishes between ambiguous situations and hard compliance failures, ensuring a package is frozen pending human review whenever a required test is missing or failed.

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)

Developer Publicly Lists Safety Flaws in His Own AI Planning Tool to Build Trust

A developer who built an open-source AI agent planning engine has voluntarily published a register of known failure modes in his own tool, arguing that transparency strengthens rather than undermines credibility. The engine previously blocked all 11 adversarial goals and 35 flawed plan variants tested, but community commenters identified critical gaps the developer had glossed over. Key weaknesses include the inability to re-audit tool outputs fetched after initial goal validation, meaning a malicious payload in an API response or database record could slip past safety gates. A second flaw allows structurally valid but semantically harmful plans to satisfy deterministic checks undetected. The developer has logged these gaps as tracked issues and outlined planned fixes, including typed tool-result provenance and capability-scoped state transitions, while acknowledging the solutions are not yet implemented.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds AI Tool That Converts Plain-English Descriptions Into Regex Patterns

A developer has launched DataBench, a tool designed to eliminate the manual step of translating plain-language requirements into regular expression syntax. Users describe what they want to match in plain English and receive a working regex pattern, a breakdown of each component, and auto-generated test cases. The tool is positioned as a complement to established platforms like regex101 rather than a direct replacement, as it targets the pattern-creation stage rather than testing and debugging. The developer openly acknowledges limitations, including narrower programming language flavor support, a smaller community pattern library, and rate-limiting on AI-generated outputs. DataBench is currently available for free at databench.dev, with usage caps applied per IP address over a rolling time window.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to enable zsh autocompletion for Composer and Symfony Console commands

A developer has shared a single zsh completion script that adds autocompletion support for both Composer and Symfony Console commands. Once installed, Composer commands like 'remove' and 'reinstall' autocomplete using packages actually listed in the project's composer.json file. The script also suggests Packagist package names sorted by popularity when running 'composer require', eliminating the need to copy-paste from the website. Installation requires just a few terminal commands to download the script and update the zsh configuration file. The author notes the workaround will become unnecessary if Composer updates its PHP compatibility policy and upgrades its symfony/console dependency.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Shared Task-State File Keeps AI Agent Work From Going Off the Rails

A developer working with an AI agent on a blog post found that small follow-up questions repeatedly pulled both parties away from the original task into nested investigations. Drawing on the concept of a call stack from programming, the author argues that a shared markdown file tracking the current task frame, working hypotheses, and return steps can preserve context across these tangents. Unlike an agent's internal to-do tracker, which may not survive session resets, this external file acts as a mutual handoff surface between the human's and agent's attention. The author notes that drift is not caused solely by the agent — human curiosity and mid-task interruptions are equally responsible for losing the original thread. Maintaining a checkpoint before any focus shift, the piece concludes, makes deep dives reversible rather than derailing.

0 comments Read more at DEV Community