Replit’s CEO says an idea from 1937 could make the org chart obsolete

Replit's CEO Amjad Masad envisions a future where AI agents, rather than traditional organizational structures, drive company operations. In a recent X post, Masad argued that AI agents can absorb and execute routine tasks within companies, suggesting that the internal coordination required by firms has remained unchanged for nearly a century.

This idea echoes Ronald Coase's 1937 question about the necessity of companies themselves, with Coase's answer being that coordination within firms is more efficient than through open markets.

Masad's concept, the "Agent Habitat," aims to create a controlled space where AI agents can work with employees and other agents while maintaining control over permissions and approvals. Replit's Agent Habitat is being implemented through Replit for Business, a platform that grants AI agents access to company systems, data, and workflows. However, the underlying technical specifications for enforcing these permissions have not yet been published.

The shift towards self-driving companies presents challenges, as current AI agent deployments often lack visibility into their access and actions. Recent research has also demonstrated that coding agents can sometimes exceed their intended scope, even when performing legitimate tasks. To mitigate these risks, Replit's Agent Habitat aims to balance access with strict permissions enforcement.

In implementing the Agent Habitat, Replit has introduced Cloud Agents, allowing employees to modify a company's codebase without needing a development environment or deep technical knowledge. This feature enables engineers to provide tested changes to the codebase, reducing the time and effort required for development. Replit's Cloud Agents also facilitate shared development environments, eliminating the need for employees to configure repositories, databases, and services individually.

Moreover, Replit is exploring the use of agents to learn from code reviews. When an engineer identifies an issue, Cloud Agents can suggest updates to shared instructions, or "skills," which guide future agents in avoiding similar mistakes. Engineers must approve these updates before they are implemented, and the feature is currently in early access.

Early testing has shown promising results, with a 51% reduction in token consumption and a dramatic decrease in task completion time from 16 minutes to just one minute for routine data tasks.

Replit's Agent Habitat extends beyond engineering, as designers and support teams have begun leveraging agents to build features and investigate customer issues. However, the effectiveness of these gains remains unverified, as Replit has not provided detailed data or separated the impact of agents from other process improvements.

The broader vision of the Agent Habitat is still in development, and Replit has not yet demonstrated agents handling company-wide tasks. While the concept holds promise, its success could be more challenging to achieve in large enterprises with legacy systems and processes.