AutoDesign evolves frozen LLM harnesses

// 1h agoOPENSOURCE RELEASE

AutoDesign evolves frozen LLM harnesses

AutoDesign is an open-source design agent that recursively improves its prompts, tools, orchestration, validation, and rendering logic around fixed language models. Its DesignHarness converts research papers into editable posters, slides, webpages, and narrated videos through iterative generation and critique.

// ANALYSIS

AutoDesign makes a compelling case that agent performance increasingly depends on the harness surrounding a model, not just the model itself.

–Meta-harness optimization searches over executable harness code using rollout traces, evaluator scores, and regression checks
–AutoDesign scores 78.32 on its 100-paper PosterBench track, beating Claude Design by 7.45 points
–The reusable DesignHarness improves average performance from 54.99 to 67.39 across seven model-agent configurations
–Local-first execution, provenance tracking, native HTML outputs, and editable artifacts make the system unusually inspectable
–The main limitation is evaluation scope: polished academic communication is a valuable testbed, but broader agent tasks will determine how general the approach really is

// TAGS

autodesignagenttool-useevaluationcode-generationopen-source

DISCOVERED

1h ago

2026-08-15

PUBLISHED

1h ago

2026-08-15

RELEVANCE

9/ 10

AUTHOR

Discover AI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

TUTORIAL54m ago

Anthropic Shares Claude Cost-Optimization Playbook

Anthropic’s new guide explains how developers can balance Claude’s cost, latency, and intelligence using prompt caching, token trimming, batch processing, effort controls, model selection, budgets, and multi-model routing. Its benchmarks show caching can cut agent-loop costs by up to 3.7x, while batching reduces eligible workloads by 50%.

TUTORIAL1h ago

OpenCode Go Powers $10 DeepSeek Stack

An X post showcases a low-cost coding setup combining OpenCode Go’s $10/month access to open models with the open-source DeepSeek Harness. The harness adds protocol handling, CLI, MCP, and agent integrations for DeepSeek V4.

OPEN SOURCE3h ago

Soup CLI Launches Local 8B Fine-Tuning

Soup CLI simplifies LLM fine-tuning into one YAML configuration and command, with layer streaming that reportedly trains Llama 3.1 8B using 3.32 GB of VRAM. It supports methods including SFT, DPO, GRPO, and KTO, plus evaluation and model export.