← All speakers

Harrison Stall

Conference affiliation: Anthropic

Harrison Stall works on infrastructure for running AI agents inside applications and automated processes. At Anthropic, his work on Claude Managed Agents addresses what happens beyond the model’s next answer: where agents execute, how developers follow their progress, and how they use external services without receiving reusable secrets. He was a member of technical staff at Anthropic at AI Engineer NYC 2026.

On this page

Moving agent work beyond the terminal

In his joint workshop with CJ Avilla, Harrison demonstrated a stock-watchlist workflow in Anthropic’s console. A lead agent organized the task and started subagents, while developers could inspect individual threads and their events as the work continued. Managed Agents provides API resources for defining agents, supplying tools and computing environments, and starting sessions. That makes it possible to trigger work from an application, schedule, or webhook without someone keeping an interactive terminal running.

The workshop introduced Flows as a way to let a lead agent plan work and coordinate subagents across phases, reducing the orchestration developers must write themselves. The watchlist example separated research on individual companies from subsequent review: each scanner could concentrate on its assigned company, and a fresh review context could examine the findings. Harrison and CJ presented that additional work as a tradeoff. More comprehensive investigation costs more, so developers need evaluations and deliberate choices about which phases warrant additional agents or stronger models. Their closing Q&A also qualified the broader vision of complex agent networks: the implementation then supported one layer of nesting.

Using credentials without exposing them

Credential isolation is a concrete part of the infrastructure Harrison and CJ explained. An agent with access to a computer may be able to read a key stored in a file or environment variable; hiding the file does not remove that risk. Their Twilio example instead used a credential vault. The execution environment received a placeholder, and an outbound proxy substituted the real credential when a request went to an approved host. Configuring both the destination and the location of the credential in the request allowed the agent to call the service without placing the reusable key in its sandbox.

This distinction between instructions and enforced restrictions also shaped their account of multi-agent safety. Separate context windows give agents distinct working contexts, but a shared filesystem does not automatically provide directory-level access controls. The workshop described tool-permission checks as a way for an application to deny particular file requests. Network restrictions could block unapproved destinations, while spending budgets could pause a session without discarding completed work. An event history and an interrupt API gave developers ways to inspect and redirect a run. Together, these mechanisms explain the practical scope of Harrison’s work: making agent execution available through an API while giving application builders specific ways to limit access, spending, and continued action.

Read the topics behind these talks

1 conference talk

Loading more talks and highlights…

References