Research library: AI agents, voice AI and web platforms

This library holds research reports that AR Labs carried out before making engineering decisions. Each one started from a concrete question: which framework to build a voice agent on, how a long-running AI agent should resume its work after a break, how to stop a multi-round LLM code review from finding more problems every round, or whether a small web app should default to Cloudflare. The reports are published here close to how they were written, with internal references removed, so that a reader can see the evidence behind our choices and reuse it. Each report links to the project or method page it informed.

The research draws on many sources for each question: vendor documentation, source code, academic papers, practitioner writing and benchmarks. Where a report scored its sources, it used the CRAAP test (currency, relevance, authority, accuracy and purpose) and checked important claims against more than one independent source. Claims that could not be confirmed are marked as such, and every report keeps its citations so a reader can check them.

Every report records the date it was carried out. The tools and models it covers change month by month, so a finding that held in February may not hold in November. Read each one as an account of the field on that date, not as current advice.

The reports are grouped by subject: voice AI, agent orchestration, context and memory, verification, interfaces, platforms and methods.