Prasanna Sankar: As agents write more code, coordination becomes the bottleneck

Prasanna Sankar describes building Vorflux to increase the amount of work a person can coordinate with coding agents. As agents run longer and more tasks in parallel, he says the bottleneck moves from writing code to managing and landing the work.

Sankar recounts his path from Microsoft to founding companies including Rippling. He says recent coding models and longer-running agents prompted him to build a system that spends more compute to save human time.

Moving work to the cloud requires an environment that can actually run and test the application. He lists dependencies, ports, authentication, secrets and seed data as setup problems; his team uses agents to help prepare a machine.

He says teams spend time iterating on plans, so Vorflux uses parallel exploration to produce a more detailed plan. Independent implementation work is then organized according to dependencies.

Testing should produce something a person can inspect without rebuilding the environment locally. Sankar describes browser tests, a video of the feature or a staging link as ways to review the result.

He describes multiple model reviewers looking at changes from different angles, followed by a merge queue that keeps branches current and lands pull requests. These are product design choices for the bottlenecks his team observed.

When more tasks run at once, Sankar says project management becomes harder: a person must remember what started, what is blocked and whether the result is useful. He frames the next product version around coordinating that work.

His proposed bots monitor signals such as customer-call follow-ups or production alerts and initiate work within a goal and limits set by a team. He distinguishes routine follow-through from the creative work of deciding what to build in the first place.

Sankar acknowledges that naive automation can flood a team with pull requests, silence symptoms rather than solve root causes, and spend money on unwanted work. His analogy is a new hire whose responsibilities expand gradually and whose work remains visible.

He gives a product example in which a bot identified a follow-up from a customer call, asked for approval, involved an engineer and prepared a fix. He says the engineer reviewed the change and he approved the customer follow-up.

His longer-term vision is a company where routine loops respond to calls, alerts and slow builds while people focus on new problems. He acknowledges that bots will make mistakes and that some responsibilities are not ready to hand over.