Augment Grew Output Per Developer 4.5× With AI Software Factory

Augment Code has published detailed metrics showing how its internal "software factory"—a coordinated system of AI agents operating across the development lifecycle—increased size-adjusted output per active developer by 4.5× between November 2025 and July 2026.

The company's approach extended automation beyond code generation into planning, code review, verification, feedback triage, and incident response. During the eight-month period, median time to merge fell from 11.2 hours to 3.1 hours, a 72% reduction, while the 14-day revert rate dropped from 1.9% to 0.4%.

Why it matters

Most AI coding tools focus narrowly on generating code. Augment's results suggest the real productivity gains come from automating the mechanical work surrounding code—review backlogs, test failures, feedback triage, incident investigation—while preserving human judgment on architecture, product decisions, and production risk. For engineering leaders, this points to a broader automation strategy than simply deploying coding assistants.

How the factory works

Augment built its system on Cosmos, its platform for running and coordinating agents. The factory connects four states of engineering work: Requirements, Ticket, Pull Request, and Production. Specialized agents handle transitions between states, integrating with Linear for tickets, GitHub for pull requests, Slack for feedback, and PagerDuty for incidents.

The company deployed agents incrementally based on where work accumulated, not in lifecycle order. When review backlogs grew in January, Augment automated code review. When Slack feedback channels consumed too much developer time, it deployed a feedback triaging agent.

By early August, the system included Project Builder for planning and design, PR Author for implementation, a fleet of review agents (Risk Analyzer, Deep Reviewer, Pair Reviewer), Verifier for runtime testing, Feedback Triager, and Incident Investigator.

Metrics and adoption timeline

Across 17,200 merged pull requests, raw PRs per active developer increased 2.7×, from 16.6 to 45.5. To account for potential changes in PR size, Augment constructed a size-adjusted output index weighting each PR by changed files and lines. That measure grew from 12.3 to 55.7 effective pull requests per developer.

The factory evolved over nine months. In November 2025, engineers gained access to Claude Opus 4.5 and used interactive IDE and CLI agents for nearly all new code. Cosmos launched internally in January 2026, with PR Author and initial review agents following in February. Code-review agents reached engineering-wide adoption by April. Incident Investigator deployed in May, Project Builder in June, and Feedback Triager and Verifier in July.

Design principles

Augment emphasized that agents handle mechanical work while humans retain ownership of consequential decisions. Engineers review design-and-architecture reports, inspect runtime evidence from automated testing, and make final merge decisions. The company found that agents needed access to the same evidence and environments as engineers—CI systems, test environments, logs, metrics, tickets, and operational runbooks—not just repository access.

The team also learned that adoption followed deployment gradually, as engineers learned to trust agent output and workflows were tuned for each team's needs.

The analysis and metrics were first reported by Augment Code in a detailed case study authored by Akshay Utture, who leads engineering for the company's AI code review agent.