Latest research / October 6, 2026
Dots vs. Muse for Developers
We gave both agents the same app prompt. Dots published through ChatGPT Sites; Muse built local apps and tried tunnels for public access. Compare the stacks, screenshots and results.
Amplifying research
Research
We study how coding agents find, understand, and use developer tools. Our reports follow the choices they make, the sources they read, and the code they produce.
Publications
12 publications
Tool choice
Dots vs. Muse for Developers
We gave Dots and Muse the same app prompt. Dots published through ChatGPT Sites and Cloudflare; Muse built local servers and tried tunnels. Compare their stacks, screenshots and results.
Tool choice
What Muse Actually Chooses: Four Different Stacks, Not Fully Deployed
Muse built servers, databases and sign-in across four different stacks. Making the apps publicly accessible remained unfinished.
Tool choice
What Dots Actually Choose: OpenAI’s Agent Already Has a Stack
Our benchmark shows OpenAI’s Dots defaults to its own platform, powered by Cloudflare, rather than separate third-party services when building apps.
Safety
Six boundaries to check when sandboxing coding agents
Lessons from building coding-agent benchmarks. Six isolation boundaries, practical checks, and what transcripts alone cannot establish.
Agent ecosystem
State of the Coding Agent Market
A view of the coding-agent market through attributed public commits and pull requests, with methods and limits attached to the counts.
Tool choice
What Kimi K3 & GLM-5.2 Actually Choose (China's AI models vs Claude Sonnet 5)
The same developer-tool questions, asked in English and Chinese, produce different choices. Compare Kimi, GLM, and Claude across the same repositories.
Tool choice
What Fable Actually Chooses
An examination of Fable’s tool choices and the tasks it solves with custom code. The study also tracks which vendors it names as future replacements.
Safety
The Security Decisions Claude Code and Codex Make
Claude Code and Codex receive six development tasks with security defaults left unspecified. The report checks the resulting code against 33 exploit tests.
Agent ecosystem
Claude Code's Leak: Every Hardcoded Vendor and Tool
An analysis of vendor references across connectors, documentation access, tool presentation, and other integration layers.
Tool choice
The Tools OpenAI Agreed to Buy
A comparison of Ruff, uv, and other Python tools across seven categories and three repositories. Read the selection counts and benchmark conditions.
Tool choice
What Codex Actually Chooses (vs Claude Code)
Codex and Claude Code answer the same prompts in the same repositories. Compare their choices across 12 categories, including custom code.
Tool choice
What Claude Code Actually Chooses
Open-ended tasks across real repositories, with no tool names in the prompts. Follow which tools are selected and when Claude builds its own solution.
Get research updates
New Amplifying research on coding agents, delivered by email.