Open-Source Agent Backboard R-CLI Tops Terminal-Bench 2.1 with 84.3% Accuracy

·1 views

Backboard R-CLI, a small open-source terminal agent, claimed the top spot on Terminal-Bench 2.1 this week, solving 75 of 89 tasks for an 84.3% accuracy score. The benchmark tests hard, real-world terminal tasks such as compiling code, debugging builds, and configuring servers, with results verified by an independent checker. Notably, R-CLI used the same Claude Opus 4.8 model available to competitors, yet outperformed the next-best Opus 4.8 result by 5.4 percentage points. The developers attribute the gains to engineering decisions like adaptive reasoning, smarter tool use, and efficient context management rather than any model advantage. All run configurations, logs, and pass/fail outcomes have been made publicly available on GitHub for independent scrutiny.

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)

AEL Open Standard Grades How Verifiable an AI Agent's Activity Record Is

A developer building an AI agent firewall has released Agent Evidence Levels (AEL), an open standard for grading how independently verifiable an AI agent's activity records are. The five-tier scale, AEL-0 through AEL-4, measures how much of an agent's record an outside party can verify and how much omission they can detect without trusting the vendor or operator. Each level adds a layer of evidence — from signed, hash-linked records at AEL-0 to counterparty-confirmed receipts at AEL-4 — with every level explicitly stating what it does not cover. The standard ships with a runnable reference checker and a conformance corpus so that grades must be demonstrated rather than merely claimed. The spec, reference checker, and conformance corpus are publicly available on GitHub, and the author intends to donate the project to a neutral governing body once it gains wider adoption.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Reflects on Lessons Learned Building Moonsu Link at Global Hackathon

A developer participated in their first global hackathon, which drew over 500 participants worldwide, submitting a project called Moonsu Link aimed at solving a real-world problem. Although the team did not win, the participant earned a badge from the DEVPost community in recognition of their effort. The experience highlighted the importance of collaboration, with teammates sharing ideas and dividing responsibilities to maintain momentum throughout the event. The developer noted that challenges such as uncertainty and technical friction were inevitable but ultimately contributed to growth and engineering maturity. Reflecting on the outcome, they concluded that completing a full product cycle and gaining experience are as valuable as taking home a prize.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Indie Developer Advocates 2-Week MVP Rule to Break the Build-in-Silence Trap

A developer behind OpenNomos has shared a personal rule: if a project cannot ship something usable within two weeks, too much is being built. Drawing from a past experience where three months of work went unused, the author argues that early users need one working feature, not a polished product. Their current project, OpenNomos Json, launched with just two tools — a Timestamp Converter and a JSON Parser — and attracted users immediately by solving real pain points. The author emphasizes that fast shipping creates a feedback loop, turning zero users into iterative improvements within weeks, unlike months-long silent builds that often launch to no audience. The key mindset shift, they argue, is resisting the urge to add features before going live and accepting that a rough but functional V1 is more valuable than a perfect unreleased product.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Open-Sources 17 Debugging Skills for Cocos Creator 3.x Engine

A developer has released a collection of 17 debugging skills for Cocos Creator 3.x, built from months of hands-on troubleshooting experience. The skills cover common pain points including gray screen diagnosis, corrupt scene recovery, CLI build failures, and TypeScript compile checks. Each skill is designed as a ready-to-run workflow that can be loaded into an AI agent to automate problem resolution. The full collection has been published as an open-source repository on GitHub under the account ChrisLamDev. Developers working with Cocos Creator can copy the skills directly into their local agent environment using a single terminal command.

0 comments Read more at DEV Community