Cursor Engineer Lauren Tan: Merged over 1,000 PRs last month, aiming for about double this month
Cursor engineer Lauren Tan stated that she merged over 1,000 PRs last month, with a target of about double this month; she attributes her productivity mainly to the cloud Agent and self-built engineering verification workflows, rather than simply generating code in bulk with models.
She has encapsulated her personal workflow into a pstack plugin, referred to internally at Cursor as "lauren-mode." Public descriptions indicate that the system includes 23 workflow skills, 21 engineering principles, 22 task scripts, and 2 dedicated sub-Agents; the main command routes requests to corresponding task processes for fixing, functionality, refactoring, troubleshooting, or performance.
The core constraint of this process is: building through does not equal task completion. Command line changes must run real commands, interface changes must follow actual user paths, database migrations must replay real inputs, performance modifications must compare traces, and storage layer changes must read back data for validation; repositories lacking validation methods must first generate local validation skills for the project.
She utilizes the cloud Agent to continue running while the computer is offline, stringing together the process of "reading work orders, planning, executing, testing, reviewing, creating PRs, checking CI, and merging after approval." Additionally, she uses Grok Bot to routinely collect failure reports from Slack, user feedback from X, and feature ideas, serving as outer input for deciding the next batch of Agent tasks.
She mentioned heavily using commands like goal, loop, and swarm to run the "Full Autopilot" script, with multiple Agents concurrently handling different tasks, and then converging results through coordinators and review mechanisms. The premise for automation is to break down small tasks into verifiable granularity within about 5 minutes, and to distill failure modes into reusable rules, rather than handing unverified large requirements directly to the model.
In terms of market mechanisms, the coding Agent reduces the marginal costs of implementation and repetitive troubleshooting, but the verification capability determines the safe parallelization limit. Teams with executable testing environments, product control permissions, observability data, and strict CI access will benefit; organizations with weak testing, reliance on manual acceptance, or difficulty reproducing online states will face higher integration, regression, and security risks due to the increase in PR numbers.
Source: Public information
ABAB AI Insight
Lauren Tan previously participated in Meta's React Compiler, which aimed to automatically optimize React applications using compiler technology, relying on strict correctness semantics, testing, and performance validation. She has transplanted similar engineering ideas to Agent collaboration at Cursor: instead of asking the model to "write like a senior engineer," the focus is on turning the environment, task boundaries, validation evidence, and failure loops into machine-executable rules. The design of pstack is therefore closer to engineering scaffolding rather than a collection of prompts.
The capital and resource pathway has shifted from "buying model calls" to "building Agent production lines." Cloud Agents handle continuous execution, Grok Bot collects external and internal signals, pstack is responsible for task routing and validation, and CI and code review manage final merge access; this shifts engineers' time from line-by-line coding to selecting tasks, defining acceptance criteria, maintaining rules, and handling exceptions. Cursor, GitHub, cloud computing providers, and observability tool vendors will gain demand from more Agent executions, builds, tests, and inference calls, while tools with only code generation capabilities are more likely to be compressed into replaceable layers.
Historically, continuous integration transformed "integrating after submission" into high-frequency builds and automated testing, while DevOps shifted deployment responsibility from operations teams to development processes; Agent workflows are a continuation of these two transformations. GitHub Actions and CI/CD automate repetitive executions, while pstack attempts to automate requirement understanding, code modification, running validation, PR generation, and partial review. The industry is still in the early stages of expansion: tools can quickly amplify the throughput of low-risk, verifiable tasks, but cross-service failures, ambiguous product requirements, and security-sensitive changes still require high-quality human judgment.
Essentially, this is about technological substitution, but what is being substituted first is not the engineers, but the waiting, switching, repetitive searching, and verifiable execution steps in engineering. The mechanism at play is that models can already generate candidate implementations, cloud environments allow multiple candidates to run simultaneously, and testing and observation tools can convert some "correctness" into determinable evidence; without this layer of validation, Agents will only translate generation speed into faster technical debt and higher review burdens.
ABAB News · Cognitive Laws
Generation determines speed, validation determines scale
The limit of automation is not model capability but acceptance capability
Humans do not need to write every line of code, but must define what correctness is