Brex: Running Agents Well Is Mostly an Infrastructure Problem

Brex's Building Autonomous Agents for Technical Tasks (Graham Fuller, April 2026) is the infrastructure report, the second of the three reports. Written around a gRPC migration across 400+ services, its one line is: "The gap between 'AI agents can do this' and 'AI agents are doing this reliably' is mostly an infrastructure problem."

Agents stalled not because the model was weak but because CI logs, review-bot comments, and stack traces never reached them. Engineers ended up relaying that feedback by hand, "an expensive solution to a plumbing problem."

What it establishes:

• One environment per task. Each task runs in its own Remote Developer Environment with the engineers' full toolchain; work enters from Slack, Linear, GitHub, or cron, and the loop closes itself until a green PR is waiting.
• Parallelization is a systems problem. The MVP was three Python scripts; running 50 agents at once is orchestration and isolation work, not prompting.
• Aim at 40%, not 100%. Chasing full automation "yields zero value" on the long tail, and Brex runs the platform with 25 of a roughly 350-person org (First Round Review, James Reggio, Sept 2025).
• Attribution and least privilege, stated early. Agent work "needs to be attributed back to an actual human and not just a bot," ahead of any compliance requirement.
• What it leaves open. No revert rates, cost figures, isolation detail, or audit record.

The guide builds on it in one environment per task, thinking in loops, and binding the agent to a human; the Brex case study sets it beside Stripe, Ramp, Spotify, and Uber.