Day 1 · Recorded 6 May 2026
How to build a company-internal background agent system
When agent use grows from one session on localhost to fleets of concurrent work, the laptop stops being the right boundary. Cole Murray turns the question into an implementation guide: who should use an internal background agent system, what architecture it needs, and what breaks first when you try to roll it out to real teams.
Cole Murray, Creator, Open Inspect
What's in this session
Most teams can see the appeal of background agents, but the first hard question is where they should run. Cole starts with the limits of localhost: one machine, scarce CPU and RAM, Git worktrees, duplicated dependencies, and non-engineers who should not have to maintain a full dev stack.
In this session, he walks through Open Inspect as a reference architecture for internal background agents: a Cloudflare control plane, Modal and Daytona sandboxes, OpenCode as the agent, and interfaces that can meet engineers, PMs, support, and SRE workflows where they already happen.
For platform leaders, the takeaway is practical sequencing. Start with scoped use cases like SRE triage, customer support, product prototypes, and self-serve analytics; make the sandbox reproducible; wire in production context carefully; then expect review capacity and adoption mechanics to become the next constraints.
Inside the recording
- 00:00 Why internal background agent systems are emerging
Cole frames the pattern behind company-built agents at Ramp, Stripe, Coinbase, and beyond. - 03:00 Local machines stop scaling
More sessions, more contributors, and more context make laptops the wrong runtime boundary. - 06:00 PMs, analysts, and prototypes as first use cases
The best internal systems expand who can turn product intent into reviewable work. - 12:00 Open Inspect's control plane and sandboxes
A Cloudflare control plane, Modal and Daytona sandboxes, and OpenCode power the reference stack. - 18:00 Deployment lessons after the first rollout
Dev environment setup, production APIs, and reproducible tools become the real integration work. - 24:00 Driving adoption and handling review bottlenecks
Start with scoped bugs, make usage visible, then expect code review to become the next constraint.
More sessions on agent infrastructure
- Building Minions: agents on a 30-million-line codebase — Alistair Gray, Stripe
- From Assisted to Delegated: Cloudflare's AI Engineering Stack — Rajesh Bhatia, Cloudflare
- Background Agents for Genomics: Operating Across Science and Cloud Infrastructure — Xiucheng Quek, Genentech, Inc