What happened
In a test of PurpleDoubleD's locally-uncensored app (AGPL-3.0-only), a coding agent was told to fix a one-line bug; the 3B model read the files and ran the failing test but stopped without editing, the 7B model looped on a non-existent src/ folder until the Loop guard halted it, and only the 14B model fixed calc.py and passed the test.
Why it matters
The app's own mechanics—file reading, diff-based change proposals, approval-gated shell execution, and loop detection—worked correctly across all models, so the failures point to model capability rather than the app itself; for practical coding-agent use, users may need to choose models at the 14B level or above.
Summaries like this, every morning.
The locally-uncensored project, which gained over 200 stars in a day on GitHub Trending, aims to bundle chat, image generation, video generation, and a coding agent into one local desktop app. This test focused solely on the coding agent's ability to fix a simple bug, setting aside image and video generation which require ComfyUI and 6–12GB of VRAM. The app itself is built with Tauri (Rust+React) and can run in a browser-based development mode, which was used for this test.
The test environment included an AMD Ryzen 7 7700 CPU, 30GiB RAM, and an NVIDIA GeForce RTX 5060 Ti 16GB GPU. The app auto-detected an existing Ollama installation and listed available models. The onboarding recommended 9B-class abliterated models as "Agent-ready," hinting that the app's creators already recognized the need for larger models for agent tasks. The results confirmed this: the 3B model stopped after seeing the test fail, the 7B model got stuck looking for a non-existent src/ directory, and only the 14B model successfully edited the file and passed the test.
One notable observation was that in Ask mode, file edits are staged as "Pending" and not applied until approval, but the agent can proceed to run shell commands before approval. This caused the 14B model's first test run to fail because the edit hadn't been applied yet, leading to an extra round of editing and testing. The author suggests that waiting for edit approval before allowing shell execution would be more natural. Overall, the app's infrastructure—including diff-based staging, approval-gated shell execution, and loop guard—worked correctly, but the coding agent's success depended heavily on model capability, with 14B being the minimum for this simple task.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Google's Gemini 4 Carbon took the #1 spot on MathArena with an expected 85.4% score, displacing OpenAI's GPT-6…
The developer behind zero-trust agent tool ACP ended its development, renamed the repo to sealbridge, and publ…
On November 10, 2026, Anthropic released Claude Sheets in beta (building spreadsheets from prompts, 144 suppor…
A practitioner writes that after enriching specs and CLAUDE.md rules, AI still produced off-intent behavior, a…
The author fed the 'desired candidate profile' text from 145 job postings to Claude, turned the extracted labe…
Researchers de-duplicating leak-site data say alleged ransomware attacks reached nearly 3,000 over the summer…
