Isabella Kai He is a member of technical staff on Anthropic’s Applied AI team, developing Claude Managed Agents infrastructure for reliable, secure, long-running AI systems. Her work addresses a fundamental production challenge: agent software must adapt as quickly as the models it orchestrates.
Before entering the industry, He founded SN Inclusion, which supported vocational training and employment for neurodivergent people, and co-founded AUesome, which developed at-home therapy resources for autistic children. She also worked with the Stanford Neurodiversity Project.
At Stanford, she studied computer science and management science and engineering, graduating with distinction in computer science. Her earlier experience included SGNL, The D. E. Shaw Group, Intuit, Bank of America, and Mooch. She joined Anthropic after college, drawn to work combining product development, research, and enterprise adoption.
Building agents for real-world execution
He’s approach to production agent infrastructure centers on four architectural concerns:
- Model-adaptive agent harnesses: Workarounds for one model generation can impair the next. Claude Sonnet 4.5’s tendency to end tasks prematurely near its context limit prompted context resets; when Claude Opus 4.5 eliminated that behavior, those resets created unnecessary latency and caching problems.
- Separating reasoning from execution: Decoupling the agent loop from sandboxed tools lets reasoning begin while containers initialize, enables recovery when execution environments fail, and supports customer-controlled private infrastructure.
- Durable sessions and recoverable context: Distinct agent, environment, and session resources preserve interaction histories, allowing interrupted systems to resume and recover context removed from an active window.
- Outcome-based agent evaluation: Developer-defined success criteria and failure conditions allow a separate grader to determine whether a task is complete and send the agent back to work when it falls short.
