|
April 30, 2026
What Stripe, Uber, and Harvey told me about background agents
Ahead of the Background Agents Summit, this issue distilled preparation calls with engineers building background-agent infrastructure at Stripe, Uber, and Harvey.
The takeaway was straightforward: the teams furthest ahead already have working systems, shared infrastructure, and clear opinions about what matters. But they are still actively working through the hardest problems around alignment, review, and how to give agents the right amount of autonomy.
Three Things That Stood Out
|
|
The future of engineering looks conversational: Slack threads were emerging as the place where humans assemble context and hand work to agents.
Agents need a real computer: teams with strong cloud dev environments were able to move faster because agents had isolated places to work.
Review is still the bottleneck: once agents can generate large volumes of PRs, review and pre-work alignment become the new constraint.
|
The Conversational Canvas
A recurring theme in the prep calls was that the real collaboration surface is no longer just the IDE. It is the conversation around the work: the Slack thread where stakeholders add context, clarify intent, and agree on what the change should accomplish before an agent starts writing code.
That matters because a PR is downstream of alignment. If the thread already captures the plan, the review burden at the end gets lighter, even if it never disappears.
Give Agents a Computer
Another shared lesson was how much existing developer infrastructure determines how quickly teams can stand up useful background agents. Uber's remote development environments and Stripe's reproducible dev boxes created places where agents could operate without stepping on anything they should not.
The broader pattern is that human-facing developer infrastructure becomes agent-facing infrastructure too. Teams with better environments, stronger isolation, and clearer workflows are the ones that can move agents into production faster.
Summit Theme
What none of them have fully solved yet
Review at scale remains unsolved. Risk-tiering helps. Routing helps. Shared context helps. But once agents can produce hundreds or thousands of PRs, human attention becomes the scarce resource again.
|
Review Is Still the Bottleneck
The issue was explicit that high-volume agentic systems shift the bottleneck instead of removing it. Code generation becomes cheap. Review capacity does not. Teams are responding with risk-tiered review, better routing, and more alignment before work begins.
That is why the summit agenda centered on context, runtime design, and security just as much as model capability. The challenge is no longer whether agents can produce code. It is whether the surrounding system can make that output useful, reviewable, and safe.
Summit recordings
If these questions are the ones your team is working through right now, this issue pointed readers to the summit as the place to hear directly from the teams operating background agents in practice.
View the Background Agents Summit recordings.
Talk soon,
Lou
Field CTO, Ona
|