Guillermo Rauch: The future of software engineering is agents talking to each other across companies
Vercel CEO Guillermo Rauch posted on October 29, 2026, that the future of software engineering is agents talking to agents, even across different companies. He said Vercel is learning from the feedback that agents send every day to improve its products, and suggested that developers add instructions to the AGENTS.md file in their projects: when there is something that can be improved, send feedback to Vercel.
The post cited an update published the same day by the official Vercel Developers account: agents can now use the submit_feedback tool through Vercel MCP to submit bug reports or feature requests in natural language, along with structured context. Rauch framed this as an early practice of cross-company agent collaboration, rather than something limited to within a single company.
Vercel’s recent public data and product direction show that a large amount of deployment activity on the platform has been driven by coding agents; the company has also built agents internally to handle tasks such as support, code review, sales leads, and data analysis. Rauch has previously emphasized that software is shifting from direct human operation to agent-native interfaces, and that companies need to open up capabilities to agents through APIs, MCP, and CLI. This feedback mechanism directly incorporates observations from user-side agents into Vercel’s product-improvement loop.
Source: public information (Guillermo Rauch’s original post, Vercel Developers announcement, and Vercel-related public interviews and data)
ABAB AI Insight
Rauch’s wording elevates agent feedback from internal tool logs to a cross-company source of product signals. Vercel had already been using internal agents such as @v to handle day-to-day operations, and had observed external agents directly triggering deployments and purchases. By opening an MCP feedback interface this time, it effectively allows external agents to feed back friction points in structured form while carrying out tasks, creating a closed loop. AGENTS.md, as a project-level instruction file, is being used as a carrier for cross-project, cross-company conventions rather than merely an internal company guideline.
This mechanism changes how feedback is collected: it no longer depends on humans filling out tickets, but instead on the agent using the product automatically submitting feedback in context. For Vercel, this shortens the path from real-world usage friction to product iteration; for users, it requires explicitly authorizing in agent instructions when to send feedback to an external company. Cross-company agent conversation is still limited by whether both sides expose the corresponding interfaces, but Vercel is proactively exposing it as a platform capability, showing that infrastructure providers are beginning to reserve channels for agent-to-agent interaction rather than serving only human developers.
Cognitive laws
If agent feedback is returned in structured form, the product-improvement cycle shifts from human tickets to real-time signals during execution.
AGENTS.md may evolve from a project convention file into an executable protocol for cross-company agent behavior.
Opening an MCP interface by an infrastructure company is equivalent to exposing its own product to continuous audit and improvement pressure from external agents.