Augment Code Claims 4.5x Developer Output Increase From Internal 'Software Factory' of AI Agents

Augment Code says an internal system of specialized AI agents increased size-adjusted software output per developer by 4.5x over nine months, while cutting median pull-request merge time by 72% and revert rates by 79%, according to a company blog post detailing what it calls a "software factory."

The company reports the following changes across its engineering organization between November 2025 and July 2026, based on an internal analysis of 17,200 merged pull requests:

• Size-adjusted output per active developer: 12.3 → 55.7 (4.5x)
• Raw pull requests per active developer: 16.6 → 45.5 (2.7x)
• Lines changed per active developer: 8,031 → 35,147 (4.4x)
• Median time to merge: 11.2 hours → 3.1 hours (72% reduction)
• 14-day revert rate: 1.9% → 0.4% (79% reduction)

Augment says the data comes from a read-only internal tool called Cosmos ROI Analyst, which pulls merged-PR activity from GitHub and computes throughput, cycle-time, and complexity-weighted output metrics. The company characterizes this as a longitudinal case study rather than a controlled experiment, noting that model improvements and evolving workflows occurred simultaneously and cannot be fully separated as causes.

What the factory does

By November 2025, according to Augment, AI was already writing close to 100% of new code at the company via interactive IDE and CLI coding agents. The claimed bottleneck then moved to the surrounding software development lifecycle: review, verification, planning, and incident response.

Augment's response was to deploy a chain of specialized agents across four workflow transitions, built on its existing agent-coordination platform, Cosmos:

• Requirements → Ticket: A "Project Builder" agent generates a codebase-grounded design doc and splits work into tickets from a feature description.
• Ticket → PR: "PR Author" and "Ticket Manager" agents draft pull requests and fix CI failures, merge conflicts, and review comments.
• PR → Prod: A "Risk Analyzer" auto-approves low-risk PRs; "Deep Reviewer" flags objective bugs; "Pair Reviewer" produces a design-and-architecture report; "Verifier" generates runtime evidence such as screenshots and logs from end-to-end testing.
• Prod → Ticket: A "Feedback Triager" and "Incident Investigator" turn product feedback and PagerDuty alerts into evidence-backed root-cause analyses and remediation tickets.

Augment says engineers retain sign-off at each transition — approving designs, reviewing auto-fixes, making final merge decisions, and approving remediation actions — rather than being removed from the loop entirely.

The company says it did not build these components in lifecycle order, instead adding automation wherever work was piling up: code-review agents shipped in February 2026, Incident Investigator in May, Project Builder in June, and Feedback Triager and Verifier in July, following the release of Claude Opus 4.5 access and a legacy code-review bot in November 2025.

Augment says a typical pull request previously required 10–15 human touchpoints — reviewers prompting agents to explain changes, authors prompting fixes for CI failures and review comments, manual test-environment setup — and that its automated review workflow cut this to three: reviewing the Pair Reviewer's report, inspecting Verifier's runtime evidence, and making the final merge call.

What this means

These figures come entirely from Augment's own internal analysis of its own engineering organization, using its own tooling to measure the results — there is no independent verification, and the company explicitly acknowledges this isn't a controlled experiment. The November 2025 baseline also coincided with the team gaining access to Claude Opus 4.5, which Augment calls the first model reliable enough for end-to-end routine engineering tasks, muddying how much of the gain is attributable to the surrounding agent infrastructure versus underlying model capability improvements over the nine-month window.

Still, the report is a useful data point on where AI-driven engineering teams say the bottleneck is heading: once code generation is largely automated, the constraint shifts to review, verification, and operational response — precisely the areas Augment targeted. Falling revert rates alongside rising output volume, if the methodology holds up, would be a meaningful signal that speed didn't come at quality's expense. Expect competitors offering agentic coding platforms — including Cognition, Replit, GitHub, and Sourcegraph — to publish similar internal metrics as they push beyond code-completion tools toward end-to-end SDLC automation.