Design Anything Keeps the Brief Intact in Large Projects: +25% Across 24 Artifacts

An agent that can draw one good logo is not yet a designer. A real project consists of dozens of linked artifacts: a logo, a color palette, packaging, social media assets, a presentation. Each one builds on the decisions made in the previous ones.

This is where agents usually break down: by the twentieth file the brief is forgotten and the colors no longer match. The team behind Design Anything has proposed a harness that solves this problem architecturally. With a fixed LLM, it satisfies 85.

0% of the brief's factual constraints with DeepSeek V4.1 Flash High and 92.5% with GPT-5.

6 Sol High. The strongest baselines on the same models trail by 15.4% and 14.

4%. On the longest tasks, the gap widens to 25.1% and 23.

0%.

The brief lives in a graph, not in the context

The idea is simple: intent should not be left inside the model's context window. The window overflows, details blur, and early decisions get pushed out by recent ones. So Design Anything moves the intent outside the model and stores it in an intent-execution graph. This graph links every output to the process step that produced it.

The mechanism has three parts:

• a task contract stores the project's constraints and the decisions made so far;
• the design options that were chosen become explicit inputs for subsequent generations;
• a deterministic controller uses the links and execution records to check the types and number of output files, find unfinished items, and steer the agent's next move.

The key word here is "deterministic." The controller does not ask the model to honestly assess whether everything is done. It counts from the records. If the brief calls for six artifacts of a given type and the graph holds five, the missing one is named and put into work. This removes the least reliable part from completeness checking: the LLM's self-report.

228 tasks and projects of up to 24 artifacts

To test the approach, the authors built DesignAnything-Bench. It contains 228 tasks across 19 design domains, including 38 project families. Each family comes in three scales: 6, 14, and 24 artifacts. This grid shows how quality drops as a project grows, rather than just comparing models on single images.

The results are telling. The longer the project, the bigger the harness's advantage: at 24 artifacts, the gain over the best baseline is 25.1% for DeepSeek V4.1 Flash High and 23.0% for GPT-5.6 Sol High. Baseline approaches degrade as the number of files grows, while the graph holds the line. In addition, with each of the two models, Design Anything leads or shares first place on constraint satisfaction in 18 of 19 domains. This suggests the gain does not rest on a couple of convenient categories.

The comparison is fair in one important respect: the LLM is held fixed. The improvement comes not from swapping the model but solely from the harness around it. The gap between the Flash model and GPT-5.6 Sol remains (85.0% versus 92.5%), so a strong base model still matters, but a weaker model with a good harness closes much of the distance.

The authors also asked people. In a blind evaluation, professional designers more often chose Design Anything's complete sets over the baselines' results on the same models. The automatic metrics on factual constraints are backed up by professional judgment.

What to take into your own agent pipelines

The main practical takeaway reaches beyond design. Any long agentic process whose output consists of many coordinated parts benefits from external state. That applies to generating marketing campaigns, documentation, interfaces, training courses, and test suites. Three techniques from the paper transfer almost unchanged:

• record constraints and decisions in an explicit contract, not in the chat history;
• pass approved artifacts to later steps as structured inputs;
• check completeness in code by comparing the types and number of outputs against the plan, rather than asking the model.

For product teams, this is also an argument in budget discussions. If a harness lifts a compact Flash model to 85% constraint satisfaction, some tasks can be kept away from the most expensive model. For studios and agencies, the conclusion is different: value shifts from a single beautiful generation to managing the project as a whole. The winner is whoever can guarantee that all 24 files from the brief are done and consistent with one another.