Developer spends $20k and 40 billion tokens building open-source AI agent harness in Dart

·10 views

A solo developer has spent approximately $20,000 across 40 billion AI tokens over three months to build >_Fa, an open-source agent harness written in Dart, with zero lines of code written by hand. The project was built using low-cost models such as Kimi K3 and GLM-5.3, routed per task to keep costs down. The resulting tool is a roughly 7 MB single binary that requires no runtime, installs in under a second, and runs across multiple platforms including mobile, web, CLI, and CI pipelines. Notably, the harness is self-building — an automated factory shipped five releases in a single day without any human involvement. The app has now received Apple App Store approval, making it possible to build mobile apps directly from a mobile device.

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)

VS Code AI Assistants in 2026: Free Tier Limits, Top Tools, and Privacy Risks

By mid-2026, AI assistants have become the most-installed category of VS Code extensions, with 41% of users activating them weekly, according to GitHub data. Tools like GitHub Copilot, Codeium, and Amazon Q dominate the market, though all free plans impose daily usage caps that most developers exhaust within two hours of focused work. Codeium has seen rapid growth, reaching 3.5 million active installs — a 190% rise in 12 months — while outperforming Copilot on Python test generation in independent benchmarks. A key privacy concern has emerged: most free-tier assistants transmit code snippets to remote servers by default, posing potential risks for developers working on proprietary codebases. Tabnine remains the primary option for fully local inference, though its free plan excludes chat and advanced features.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How vLLM Solves GPU Memory Bottlenecks in Large Language Model Inference

Modern LLM text generation is primarily a memory bandwidth problem, not a compute problem, with GPUs often running at low utilization while shuffling model weights and KV cache data during the decode phase. Standard PyTorch inference pipelines suffer from memory fragmentation, static batching inefficiencies, and high driver overhead when handling multiple concurrent users. vLLM addresses these limitations through two core innovations: PagedAttention, which manages KV cache memory more efficiently, and continuous iteration-level batching, which schedules new requests as soon as any sequence completes. Benchmarks for the article were conducted on dedicated NVIDIA H100 and H200 GPU clusters, using models such as Qwen3.6-27B as practical examples. The piece provides a technical deep dive into why naive inference stacks fail at scale and how vLLM's operating-systems-inspired engineering overcomes those constraints.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google, OpenAI and X Move Ads Into AI Chat Interfaces Across Platforms

Major tech platforms including Google, OpenAI, and X are testing advertising formats embedded directly within conversational AI interfaces rather than in traditional placements. Google announced Gemini-powered ad formats at its Marketing Live 2026 event, designed to appear within AI-driven search and discovery experiences. OpenAI has made ChatGPT Ads available in several regions, placing ads within and alongside chat threads based on conversation topics and user interactions. X has separately experimented with contextual ad units integrated into social post conversations and discussion flows. The shared trend signals a cross-platform shift toward dialogue-aware advertising, where relevance depends on matching a user's conversational intent rather than interrupting their attention.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dart vs Freezed: Benchmark Tests Four Ways to Write a Data Class

A developer benchmarked four approaches to writing a data class in Dart — Freezed 4, Equatable, pure Dart 3.13, and Freezed 4 with primary constructors — using a shared suite of 16 tests on a single Note model. The comparison measured lines written, lines generated, build time, and how many places need updating when a new field is added. Dart 3.13, which made primary constructors stable on 12 August 2025, has been gradually closing the gaps that originally made Freezed necessary back in 2020. No single approach outperformed the others across all criteria, with each breaking down in different scenarios. The findings point toward a recommended default approach, with identified exceptions where that default is not the best fit.

0 comments Read more at DEV Community