AI is making fresh moves in coding, knowledge retrieval, and data workflows. Microsoft, Meta, and several universities have rolled out new tools that push boundaries.
Microsoft and its academic partners created SkillOpt. It’s a text-space optimizer that transfers skills between large language models without retraining the whole model. The team includes researchers from Shanghai Jiao Tong University, Tongji University, and Fudan University.
SkillOpt works by training a single natural-language skill document. The target model itself stays frozen. An optimizer reads scored rollouts and suggests bounded edits like add, delete, or replace. This means skills can move from one model to another with minimal fuss.
Tests show promising results. For example, a skill trained on GPT-5.4 kept 82% of its effectiveness when moved to GPT-5.4-mini. But the smallest variant, GPT-5.4-nano, retained only 16%. Cross-model transfer within the GPT-5.4 family showed mixed retention but proved the concept.
SkillOpt also works across different code assistants. Transferring the SpreadsheetBench skill from Codex to Claude Code raised performance from 22.1 to 81.8, surpassing the 80.4 score from training Claude Code from scratch. However, the LiveMath skill transfer from Claude Code back to Codex held only 10% of its original gain.
GraphRAG Boosts Complex Question Answering
Microsoft also developed GraphRAG, a new retrieval method. It improves global understanding, multi-hop retrieval, and complex reasoning compared to vanilla vector retrieval-augmented generation (RAG).
GraphRAG uses a graph-based approach that better “makes sense of the whole corpus.” In head-to-head tests over million-token datasets, GraphRAG won 72 to 83% of comprehensiveness comparisons and 62 to 82% for diversity against vector RAG. It’s especially strong on tough multi-hop questions, winning two out of three times or better.
GraphRAG’s summaries use up to 97% fewer tokens than processing the source text directly. This saves time and processing power. Recall@5 on multi-hop question answering jumped from 73.4% with vector RAG to 87.8% using GraphRAG.
A 2025 study by Michigan State and Meta found no clear winner between RAG and GraphRAG. They work best together. For single-hop factual lookups, plain RAG scored higher (F1 64.8 vs. 63.0). For multi-hop reasoning, GraphRAG outperformed with 70.3% versus 67.0% accuracy. GraphRAG also boosted complex reasoning and contextual summarization by 10 and 13 points.
Building the graph is costly, around $48 per index for a moderate corpus. Plus, evaluation with LLM judges shows biases like position and length preference. Still, GraphRAG offers clear advantages on large, complex datasets.
Meta’s Muse Code and DataFlow-Harness Transform Developer Tools
Meta launched Muse Code, a new coding agent powered by Muse Spark 1.2. It aims to handle tough software engineering tasks. This includes writing code, planning changes, and validating results. Muse Code can manage multiple sub-agents, delegating work for better efficiency.
Muse Spark 1.2 improves code generation, debugging, codebase understanding, and full developer workflows. The cost is competitive: $1.25 per million input tokens, 15 cents per million cached inputs, and $4.25 per million output tokens. Meta offers a discounted contributor tier at 10 cents per million input tokens and 20 cents per million output tokens. In contrast, Claude Code and Codex subscriptions range from $20 to $100 per month.
Meta’s CEO Mark Zuckerberg said in April that developer tools aren’t the company’s main focus. Still, he isn’t opposed to offering coding tools. Meta claims Muse Code is “one of the most affordable coding agents on the market, available globally.”
Alongside Muse Code, Meta introduced DataFlow-Harness. This open-source framework guides LLM agents to build structured, visual data-processing workflows step-by-step. It represents pipelines as directed acyclic graphs (DAGs) and lets engineers inspect, test, and govern each stage.
DataFlow-Harness uses four main parts: the Data Pipeline Backend (the DAG), DataFlow-WebUI for natural language and visual editing, MCP Tools Layer providing live operator registry and pipeline state, and DataFlow-Skills injecting domain knowledge as markdown files.
Tested on 12 data-engineering tasks using Claude Opus 4.7, DataFlow-Harness achieved a 93.3% pass rate. This is within 0.9 percentage points of a context-aware Claude Code baseline at 94.2%. It cuts API costs by 72.5% and speeds up responses by 49.9%. It’s also 17.6% faster than the context-aware version.
DataFlow-Harness excels in building data pipelines for tasks like question-answer generation and math data cleaning. In a textbook-to-visual QA extraction, it hit 97.2% precision and 87.3% coverage. This approach applies typed, incremental changes to a persistent DAG, making workflows transparent and manageable.
These new tools show how AI is evolving beyond single tasks. Skill transfer, smarter retrieval, affordable coding agents, and structured data workflows are shaping the future of AI-powered development.
Based on
- Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude Code Harnesses — marktechpost.com
- Stop graphing everything: When GraphRAG actually beats vector RAG | VentureBeat — venturebeat.com
- Meta to take on Anthropic’s Claude and OpenAI’s Codex with new coding agent | Business Insider Africa — africa.businessinsider.com
- Meta Introduces Muse Code, Its Take On A Coding Agent — engadget.com
- Structured AI data pipelines score 10.9 points below free-form code — DataFlow-Harness closes the gap | VentureBeat — venturebeat.com
