Mastra's Factory Agents Merged 238 PRs Last Week, Humans Merged 630
On this page 6 sections
TL;DR
Mastra, the TypeScript agent framework whose @mastra/core package averages 3.7M npm downloads a week, shipped a major update to Factory on 18 November. Factory is an open-source software factory: an agentic kanban board you @-tag in Slack, backed by cloud sandboxes and live app previews. The team says Factory agents opened and merged 238 PRs across the mastra-ai org last week, alongside 630 merged by humans, and the company published its own before/after logs: median triage fell from 6.9 hours to 11 minutes, and merged PRs per engineer per week rose 78 percent.
What Factory actually is
A web app plus a Slack bot that walks work items through columns: intake, triage, plan, build, review, done. Underneath sit Mastra's own primitives: durable agents for long-running sessions, conversational memory, a markdown wiki with source citations, and cloud sandboxes where the coding agent does the actual work. Picture a kanban board where the cards work their own tickets: tag the board in a thread, an agent picks the card up, investigates, proposes a plan, codes in a sandbox, and opens the PR for you to review.
Intake can come from GitHub issues, Linear, Jira, GitLab, or incidents from Incident.io. Each stage is a customizable skill with its own model choice, guardrails, and human approval gates; Mastra runs low-risk items hands-off and holds risky ones for review. The whole thing scaffolds with npm create factory, lives in the template-factory repo under Apache-2.0, and needs a Postgres database with pgvector. If you want nothing to do with Mastra's hosted platform, npm create factory -- --no-platform scaffolds the self-hosted template.
The numbers Mastra published
The blog post is a vendor measuring itself, which is exactly why it reads well: the authors open with "Factory claims deserve skepticism" and then publish four weeks of dated, scoped logs. Scope matters: the numbers cover the mastra-ai/mastra monorepo, not customer projects, and they are self-reported.
Merged PRs per engineer per week went from 9.5 before Factory (May 31 to July 25) to 16.9 over the last four complete weeks (October 11 to November 7), of which 12.4 were human-authored and 4.5 came from Factory agents. Triage collapsed first: a new work item sat in triage for a median of 6.9 hours in August and 11 minutes in November, with the agent handling 91 percent of November triage. Issues closed by a merged PR now take 28.2 hours end to end on Factory versus 56.4 hours before, and once a Factory item is kicked off, the median time to an open PR is 76 minutes.
Adoption is the part most vendors would have blurred. Through early October, no more than three engineers ran Factory sessions in a given week. In each of the last three complete weeks, 19 to 24 did.
The post even declines the obvious brag:
"No, we won't be announcing today that engineers at Mastra are opening 1000 PRs per week, because they don't want that, users don't want that, and no one benefits except model labs."
What shipped on 18 November
App previews give every change a live URL to the real running app instead of a screenshot, and the environments persist across sessions. Cloud workspaces give each session a microVM (E2B by default on Mastra's platform), with customizable Docker image, CPU, and memory; their pitch for running ten in parallel is "your fans never spin up." Slack control went ambient, meaning a tagged factory decides whether a reply is warranted, "a highly capable member of your team, not a simpleton assistant" in their words. Model policy can be personal, hybrid, or org-level, on anything OpenAI-compatible plus gateways including the Mastra Gateway.
The interesting primitive is Signals, Mastra's mechanism for a session to listen to upstream events and steer the turn when they change. Factory uses it to watch PR review threads: when CodeRabbit posts comments on a work item's PR, the owning agent addresses them without anyone asking.
Agent identity is configurable, a detail that matters if you try this: by default PRs come from a GitHub App called Mastra Platform, but a personal access token makes the agent loop open PRs on behalf of the user who started the session.
The wrinkle: they closed the repo
In the Show HN thread, the team says it has closed Mastra's repos to outside contributors and now asks for detailed issue reproductions instead. The stated reason: PRs of marginal value have skyrocketed at the same time as agents made them cheap to fire off. sindresorhus closed his repos for the same reason. The economics that let one org merge 868 PRs in a week also flooded its contribution queue, so the queue went away.
The caveats are real: the numbers are self-reported and measured on the vendor's own monorepo, Factory is explicitly not a 1.0 ("we'll continue to evolve, improve, and harden it"), and its happiest customer sells factories. The implementation lives in the mastra monorepo under mastracode, with the standalone template still thin. The counterweight is that the code is Apache-2.0, the logs have dates and scopes, and a trial costs you one Postgres database.
Key Takeaways
- Mastra Factory is an open-source (Apache-2.0) agentic software factory: kanban stages, Slack control, cloud sandboxes, live app previews, and any OpenAI-compatible model.
- Last week Factory agents opened and merged 238 PRs across the mastra-ai org; humans merged 630 the same week.
- Mastra's published internal logs: median triage 6.9 hours to 11 minutes (91 percent agent-done), issue-to-merged 56.4 hours to 28.2 hours, merged PRs per engineer up 78 percent.
- Weekly adoption went from at most 3 engineers during beta to 19-24 in each of the last three complete weeks.
- Signals let a work item react to PR reviews on its own; agents can PR under their own GitHub App identity or yours.
- Skepticism applies: vendor-measured, on the vendor's monorepo, pre-1.0. The code and the method are public, so you can check.
Sources: Mastra blog, Show HN thread, template-factory on GitHub, Factory docs, npm registry