Latest research / September 25, 2026
Six boundaries to check when sandboxing coding agents
A fresh checkout can still share history, memory, and host services. What building coding-agent benchmarks taught us to check.
Improve how AI agents find, understand, and use your product.
See what agents choose, where they get stuck, and what to improve in your docs, examples, and tooling.
Record what coding agents do. See what to improve.
From the researchers behind the original What Claude Code Chooses study.
Two products
See what agents choose.
01 / Choices, findings, playbooks
Benchmarking
Measure when coding agents choose your product, what they choose instead, and why. Get findings and prioritized playbooks for your team.
- Product selection and competitive benchmarks
- Sources and evidence behind each finding
- Playbooks for docs, positioning, product features, and developer experience
Test what helps them succeed.
02 / Changes, tests, results
Ground Control
Make a change to your docs, skills, or tooling. Rerun the same task to see what gets better or worse, then iterate.
- Installation, authentication, and real workflows
- Recorded runs with inspectable tool calls
- Before-and-after tests of docs and tooling
More from our research
Tool choice and security.
Our original study / February 2026
We pioneered research on what coding agents choose.
Our original What Claude Code Chooses study put numbers behind agents’ developer-tool decisions. Today, we help developer tool teams measure and improve how coding agents find, understand, and use their products.
Read the original studySafety / April 2026
What security defaults do agents choose?
Six development tasks leave the security defaults unspecified. The report checks the resulting code against 33 exploit tests.
Read studyGet research updates
New Amplifying research on coding agents, delivered by email.
See Amplifying in action.
Tell us what you’re working on. We’ll show you where Amplifying can help.