Común: How a 13-person engineering team tripled its output

Común builds financial services around how Latino immigrants move money. Many US financial institutions make the process difficult: sending a remittance home in cash carries a steep fee, and opening an account at incumbent banks often requires a branch visit and documents built around a US life. Común designed its platform to remove those barriers. Customers can open an account with more than 100 forms of ID from Latin America, send payments to 17 Latin American countries with direct-to-bank delivery and cash pickup, deposit cash at retail stores and ATMs across the US, and use ComúnPay to reach anyone in the US, including recipients outside Común.

Común's roadmap is bigger than what a 13-person engineering team can pursue at once. The team wanted to launch a WhatsApp channel quickly, overhaul disputes, make a major push into stablecoins, and build an AI assistant into the app and web product.

The road to background agents

Two realizations convinced Común that the future was background agents. First, engineers were spending their days guiding local agents through work by asking them to address PR comments, providing follow-up context, and keeping track of each task. The overhead limited how many agents each engineer could run. Second, every engineer had been building AI workflows in isolation on their own machine, including personal memory, local automations, and product context. A central platform would make each engineer's improvements available to everyone at once.

Getting there took a few attempts. Común first tried Devin about a year earlier, when the product was still early, and it gained no traction internally. The team then attempted to build its own background agent, but decided the effort was better spent on its product and on the platform work that makes agents effective: a custom MCP server, faster CI, an end-to-end testing setup with seeded data, an Android emulator through Devin, custom lint rules, and "micro reviewers," which comprise roughly 20 AI checks on every PR that enforce the specific patterns Común wants in its codebase.

When Común tried Devin again, Devin could own a task end to end: understand the codebase, write the code, run and test the application, check its own work, and submit a PR for human review.

"Using Devin felt like a magical experience in terms of being able to figure out what engineers need from an end-to-end testing perspective, and get a PR to the finish line to the point where it's ready for a human to review."

— Luis, Head of Engineering & Data

Shipping new products

Común put Devin to work on new product projects first and built an internal orchestration tool called Swarm. Swarm treats Linear projects as a directed acyclic graph (DAG) of tickets. An engineer scopes a set of tickets, sometimes 20 at a time, chooses which ones to delegate to Devin based on the DAG, and sends them to run in the background while tracking every session in one view.

Because sessions keep running in the background, engineers can follow up with Devin from their phones and leave agents working after closing their machines. On each PR, Devin and the micro reviewers iterate until the work is ready while a human focuses on final review.

In the four weeks after usage took off, the backend team went from an average of 50 PRs per week to 170, a 3x increase, while delivering the WhatsApp launch and the disputes overhaul. The share of backend PRs created by Devin rose from 18% to 82%.

The stablecoins team is structuring its entire project as well-scoped tasks designed to be executed by a fleet of Devin agents when the plan is ready. Común expects to compress the original timeline by a factor of 2–3x, from 10 weeks to 4 weeks.

Handling on-call and triage

Común gave Devin the same context available to its engineers: databases excluding personally identifiable information, Datadog logs, feature gates, and data from external vendors through custom API integrations. Devin automatically picks up pages from Slack, investigates, and often arrives with a fix ready. When an alert is noise, Devin tells the engineer that no action is needed. Support agents' escalations get the same treatment.

The weekly "goalie" on each team used to spend the entire rotation on pages and escalations. Goalies can now work on foundational fixes. Lower-priority tech debt and small performance improvements also get done because anyone can send a Devin agent from Slack to investigate and fix them.

Expanding beyond engineering

Analytics gave Devin access to its own dbt environment, turning Devin into a data platform that growth, operations, and product can use directly. Product managers who find an issue ask Devin to investigate and push a fix, with engineers validating and merging the work. Data science is exploring Devin for training machine learning models and running daily workflows, and Devin is close to being able to run and test Común's mobile app end to end.

Building an engineering team around agent orchestration

Devin lets Común pursue more of its ambitions as a lean team. Stablecoins and the in-app AI assistant are both underway, and hiring now focuses on engineers with strong judgment and system design skills who can orchestrate many agents across infrastructure, backend, and frontend.

"Every small update we ship—to our Devin setup, to our verification loop, to the data context it has access to—saves hours and hours of investigation, code review, or execution time. As a company, we're trying to figure out what the horizon is before we solve it with more people, more process, more structure."

— Luis, Head of Engineering & Data