▶ Watch ↗16:43
Conference affiliation: SoFi · 2026
Ishween Kaur builds agent systems and tests what happens when they can spend money. Her work spans multilingual reasoning at Salesforce and adversarial experiments with x402 payments: whether an agent authorizes the intended purchase, pays the intended seller and stops buying when its task is complete. She was credited with SoFi at AI Engineer NYC, and her speaker profile describes her work there as senior software engineering in crypto and AI.
Ishween joined Salesforce in 2020 and worked across enterprise software teams, including MuleSoft, before moving into conversational AI and agents. Her career account traces that period from Hyderabad to San Francisco.
On Salesforce’s Agentforce Agentic Reasoning team, she worked on a problem that language fluency alone could not solve. A planner, retrieval service and action could each function individually while producing an interaction that unexpectedly switched languages. Leaving localization to individual services created conflicting fallback behavior, especially when a user changed languages mid-conversation.
The team centralized Localization Context: shared language state established before reasoning and updated by the planner. Downstream components consumed that state instead of independently choosing a language. The team selected Lingua for language detection after evaluating alternatives, including an additional model invocation. In Ishween’s account of the architecture, keeping an interaction in French requires coordinating the components around the model as well as generating fluent French.
Ishween’s independent payment experiments examine which purchasing decisions an agent should be permitted to make for itself. She tested x402, a protocol for payments over HTTP, against deliberately adversarial sellers. Her AI Engineer demonstration showed a buyer paying fifty cents for a service advertised at five cents. Another experiment exposed a gap between payment verification and settlement: several simultaneous requests could pass verification even though only one payment ultimately settled. These were test-environment failures, not reported losses in deployed financial systems.
Her written account of the experiments develops two practical decisions:
These experiments distinguish a valid payment from an authorized purchase and a completed exchange. Price, recipient, settlement and delivery each require their own checks; a successful signature does not establish all four.
Ishween also serves as communications lead for Women in AI and teaches Agentic System Design. Her CORE framework organizes that teaching around clarifying the task, orchestrating execution, remembering context, and evaluating and hardening the system. It gives software engineers moving into AI a way to examine how a model’s task, execution, memory and permissions fit together—the same surrounding machinery that determines whether an agent maintains a chosen language or makes only an approved purchase.
▶ Watch ↗16:43
Loading more talks and highlights…