Developer Builds 4-Agent AI Research System FORGE in Three Days at Home

·2 views

A developer built FORGE, a locally hosted multi-agent AI system designed to answer research questions using a pipeline of four agents: a planner, researcher, writer, and reviewer. The project took three days, with most of the time spent identifying parts of the app that appeared functional but were actually simulated or broken. Early builds used a fake simulator to populate the dashboard, while a key architectural decision — logging all activity as immutable events rather than updatable records — allowed live and replay views to share the same code. When real AI models replaced the simulator, issues emerged including IPv6 timeout failures, agents exhausting output budgets before answering, and runaway research loops, all of which were fixed with targeted patches. The final system runs on affordable models like GLM-5.3 Flash via OpenRouter and uses free tools such as DuckDuckGo search and open-source page-scraping libraries.

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)

AI Agents Cannot Distinguish Safe Actions from Catastrophic Ones, Experts Warn

AI agents operating on user interfaces treat all clickable actions as equally weighted, unable to differentiate between harmless tasks like downloading a report and irreversible ones like deleting a database. Unlike humans, agents lack instinctive hesitation or consequence-awareness, acting purely on learned patterns. This creates serious risk when agents are given unsupervised access to tools and dashboards, particularly during off-hours with no human oversight. Developers argue that permissions and approval workflows must be built as structural safeguards, not optional add-ons, defining which actions agents can perform autonomously and which require human sign-off. The core fix, experts say, lies not in improving agent judgment but in designing stronger system-level guardrails around agent access.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Bug Report: Registry Rule Had Two Sources But Only One Enforcement Step

A software registry system contained a rule defined across two carriers — a registry file and a submission template — but only one of the two had a step to enforce it. When a registered instance's runtime ID changed, the registry's Instance value went uncorrected because the automated rotation pass only checked the editor's own row, not all registered rows. The gap was discovered when a pull request updated a descriptive machine column without correcting the binding Instance value, which is the field that rules actually resolve. A repair was applied across three sites in two carriers, extending the collector step to cover every registered row and clarifying what qualifies as a valid registration act. The case also highlighted that a diff's size does not indicate its significance — the unmerged PR changed real values, but in a column no rule reads.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Open-Source AI Pipeline to Dub Anime and Movies Using Python

A developer named Sumit884 has released dub_movie, an open-source Python project that automates video dubbing by combining speech recognition, translation, and text-to-speech generation. The pipeline uses tools including Faster-Whisper for speech detection, Edge TTS for voice synthesis, and FFmpeg for final video rendering. It is primarily designed for Japanese anime and movies with SRT subtitle files, translating dialogue into other languages such as Hindi while preserving character-specific pitch and tone. The system breaks dubbing into separate stages — including speaker identification, audio separation, and timeline synchronization — to ensure translated speech aligns with original dialogue timing. The project requires Python 3.10 or later and supports optional GPU acceleration for faster local translation and processing.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Coding Horror Blog Reflects on the Value of Helping Others

A post on the long-running Coding Horror blog raises questions about the importance of mutual help and human connection. The piece appears to explore what society risks losing when people stop supporting one another. It was shared on Hacker News, where it gathered 22 points and prompted at least one comment. The blog, authored by developer Jeff Atwood, has historically addressed both technical and broader human themes in software culture.

0 comments Read more at Hacker News