Vorflux founder Prasanna Sankar describes how coding agents push bottlenecks beyond code generation. This is his account of Vorflux’s product and a customer example, not a measured comparison with other platforms.
Make cloud work testable
- Give long-running agents the same dependencies, authentication, seed data and services a developer uses locally; a cloud sandbox that runs only one service cannot prove the full change works.
- Spend effort on parallel exploration and planning, then assign independent pieces through a dependency graph.
- Run browser and end-to-end checks near the cloud workspace. Share a recording or staging link so a person can inspect the result rather than trusting a summary.
- Use review from different angles and a merge queue to keep simultaneous PRs compatible with the current base branch. These controls do not eliminate human review.
Starting the right work is the next problem
- Sankar says larger numbers of concurrent agent tasks shifted his attention from coding to remembering, prioritizing and coordinating work.
- Vorflux’s proposed always-on bots watch signals such as alerts and customer calls. In his example, a bot found a bug in a call transcript, requested permission, involved an engineer and drafted a customer follow-up after the reviewed fix shipped.
- Unrestricted event-triggered agents can flood teams with PRs and consume budgets. Give each bot a goal, named owner, visible work, spending limits and escalating permissions only after evidence of reliable outcomes.
- His “zero-employee company” is an aspirational north star; the reported example still needed human approvals and a responsible engineer.
“Nobody hires a person and says solve everything.” — Prasanna Sankar