The AI Builder Podcast
How OpenHands Runs AI Engineering Agents at Scale with Robert Brennan
OpenHands CEO Robert Brennan on running fleets of AI coding agents: persistent cloud environments, credential isolation, and evaluation that decides when work needs a human.
What you’ll learn
- Agent Canvas manages multiple agents from the browser, each running in its own persistent cloud environment.
- Work is triggered from GitHub, Slack, Jira, Linear or a schedule — not typed into a chat box.
- Role-based access control and credential isolation are what make an agent fleet safe to share with a team.
- Benchmark models against your own tasks: acceptance-based evaluation tells you when a cheaper model is good enough.
- The next productivity leap: work finds the right agent, gets done and evaluated, and reaches a human only when judgment is required.
About this episode
OpenHands began as an open-source coding agent. It has grown into something more ambitious: model-agnostic infrastructure for engineering teams that treat agents as part of the team — the layer that runs them, secures them, and decides when their work is ready.
Co-founder and CEO Robert Brennan walks through how that works in practice. Agent Canvas manages multiple agents from the browser, each in a persistent cloud environment, with work triggered from GitHub, Slack, Jira, Linear or on a schedule. Keeping a fleet safe means role-based access control and credential isolation; keeping it affordable means benchmarking models against your own tasks and using acceptance-based evaluation to know when a lower-cost model is good enough.
Brennan's core argument is the destination: work automatically finds the right agent, gets completed and evaluated, and reaches a human only when judgment is required. A practical guide for anyone designing a secure agent fleet or automating a software-development workflow.
Questions this episode answers
What is OpenHands?
OpenHands started as an open-source coding agent and has grown into model-agnostic infrastructure for engineering teams — the layer that runs fleets of AI agents in persistent cloud environments, secures them, and evaluates their work.
How do teams trigger work for AI coding agents?
Instead of typing into a chat box, work is triggered from where it already lives: GitHub, Slack, Jira, Linear, or on a schedule. Agent Canvas then manages the running agents from the browser.
How do you keep a fleet of AI agents secure and affordable?
Security comes from role-based access control and credential isolation. Cost control comes from benchmarking models on your own tasks and using acceptance-based evaluation to decide when a cheaper model is good enough.
