Merge Is the Bottleneck

Prasanna Sankar, Vorflux20:06 · Nov 2026 · 1,720 views
Thumbnail for Merge Is the Bottleneck Watch on YouTube
TL;DR
  1. 1

    Long-running cloud agents move the bottleneck from writing code to planning, testing, reviewing, merging, and project management.

  2. 2

    Vorflux's Autobots start work from signals such as customer transcripts, production alerts, and CI events, then coordinate people and agents around an owned task.

  3. 3

    Automation needs budgets, visible work, ownership, and increasing responsibility because indiscriminate agents create noisy alerts, unwanted pull requests, and wasted tokens.

Summary

Prasanna Sankar describes how coding agents changed from IDE autocomplete to cloud sessions that can work for hours. That shift exposed a series of practical bottlenecks: reproducing the developer's environment, spending more inference time on planning, splitting dependent tasks, testing through browsers and recordings, reviewing with several agents, and keeping pull requests mergeable. Once those systems allowed more parallel work, project management became the limiting factor. Vorflux's next step is a set of persistent Autobots that start work from existing company signals. A customer call can create a fix, involve the right engineer, merge the pull request, and draft the customer response. Sankar is direct about the risks. Unbounded automation produces alert floods, unwanted pull requests, silent failures, and large token bills. His proposed model gives bots goals, owners, channels, budgets, and visible progress, then expands their responsibility as their work proves reliable. He expects people to retain creative judgment while bots handle well-specified follow-through.

Key ideas
03:46

Cloud agents need a working copy of the developer's environment

Sankar says agents that run for hours belong in the cloud, but most cloud agents run blind or start only the service being changed. Good testing depends on the other services, secrets, authentication, ports, and seed data that exist on a local machine. Vorflux uses subagents to set up a raw EC2 instance and reproduce that environment before work begins. The system spends tokens and money up front on machine setup. Sankar presents this as a condition for moving development from a desktop to the cloud, rather than as a minor provisioning detail.

05:10

Planning deserves more inference time because human time is spent there

After the environment problem, Sankar found that planning consumed much of his time. Earlier planning agents returned quickly, which led to repeated back-and-forth with the user. Vorflux instead launches parallel exploration subagents. They investigate the problem, build an architecture, and produce a detailed plan of plans before generating a summary. Once the plan exists, it includes a dependency graph. Subagents can then take independent pieces of work, while dependent tasks wait for the work they need.

06:41

Testing has to produce evidence without sending work back to a laptop

Sankar treats local testing as another barrier to cloud development. Vorflux runs browser and end-to-end tests on the machine where the agent works. The result gives the user concrete evidence, such as a recording that shows the feature working or a staging link that can be clicked and inspected. Because the application is already hosted on that machine, the user can interact with it there. This removes the need to pull the code back to a local host just to test the change.

07:44

Review and merge need their own agents when pull requests multiply

Code review becomes a bottleneck when agents open many pull requests in parallel. Sankar says teams can ask several models to review different aspects of a change, such as building with Claude and asking GPT to review. Vorflux makes this part of a customizable harness. Merging needs separate machinery. Its merge queue continually rebases pull requests, keeps them mergeable with the main branch, and lands them automatically. These pieces allow longer-running cloud agents to continue without leaving people with an unmanageable queue.

08:48

More parallel coding moves the limiting factor into project management

Desktop coding might support three to five parallel tasks, according to Sankar, while long-running cloud agents can support 20 or 30. That increase changes the problem. The difficult part becomes deciding what should start, tracking progress, and making sure useful work actually lands. Sankar says his own context becomes difficult to manage once more than 10 tasks run in parallel. He connects this to model design as well: recent systems use subagents for heavy work and let a main agent orchestrate fresh context windows.

10:39

Existing company signals can start routine work without manual task entry

Sankar says many tasks already contain their own intent. A customer call can leave to-dos in a transcript, a production alert can indicate what needs investigation, and an alert or codebase can provide enough information to begin. Treating agents only as a queue means work moves as fast as the person feeding it. Vorflux's Autobots instead watch these signals, take ownership, and collaborate with the team. Sankar draws a boundary around creative zero-to-one work, which still needs human taste and judgment. Bots handle the well-specified follow-through.

12:58

Unbounded automation fails through noise, wasted work, and poor visibility

Sankar describes teams telling automation to solve every Sentry issue. The result can be a flood of pull requests nobody has time to merge, constant notifications, alerts that get silenced without fixing the cause, and token costs from work nobody wanted. He compares a useful bot to a new hire whose responsibility grows as their work holds up. The bot should work within a budget, show what it is doing, and receive important tasks gradually. Vorflux gives each Autobot a goal and a home channel, with people able to approve or reject its work.

14:17

A customer-call bot can carry a fix from discovery through follow-up

Sankar gives a customer example. During a beta rollout, a bot failed to respond in Slack. After the call, an Autobot read the transcript, found the to-do, and used the screen recording to locate the relevant moment. It asked Sankar for permission to investigate, tagged the engineer who owned the fix, produced a pull request, and waited for approval. After the engineer approved it, the change was merged and deployed. The bot closed the item and drafted a customer follow-up, which Sankar approved before it was sent.

"We climb the leverage one rung at a time as a new bottleneck shows up."18:35
Who should watch
  • You are moving coding agents from a laptop into cloud environments and need to solve setup, testing, review, and merge problems.
  • Your team can generate more pull requests than people can track, and you need a model for ownership, budgets, and approval.
  • You are designing bots that should react to customer calls, alerts, CI events, or project signals without creating a noisy stream of unwanted work.