Retool begins early access to centralized agent harness

Retool announced early access to a centralized agent harness on Oct. 6, 2026. It gives employees individual agents they can use in Slack, Teams and Google Chat, while company administrators manage which tools, data and models are available. Retool said it is beginning to offer the product to its next hundred customers over the coming weeks. The company also said it has used the system internally for four months, with new hires receiving it by default.

Retool frames the product as an alternative to relying on agents running separately on employee computers. The company says its engineers use coding agents locally, where a development sandbox lets them test changes and undo mistakes. But it argues that agents connected to live business systems pose different risks, such as sending an incorrect message that cannot simply be rolled back. Its proposed approach is to manage business agents through shared company infrastructure.

At the center is a common policy and execution layer between agents and company systems. Tool requests go through a gateway where permissions and policies can be applied, and actions can be associated with the agent that made them. The system can run centrally or be deployed on-premises in a company’s own virtual private cloud. Each employee can have one or more agents, with separate filesystems, memory and accumulated context.

Administrators can choose which frontier or open-source models different users may access, and route work according to desired quality and cost. Retool says policies can inspect tool-call arguments, so allowing an agent to use a tool need not grant unrestricted use. Those checks operate outside the model, according to the company, rather than depending only on instructions the agent is expected to follow. Retool also says it excludes employee laptop access, browser automation and desktop control; agents instead run in a hosted sandbox and use governed tools to reach business systems.

Actions, inputs and outputs are logged by person, model and job, and the company says this lets administrators trace token spending to the work it supported. Agents can also perform scheduled tasks, which users can create by describing recurring work. For AI builders, the design combines centrally managed model selection, tool permissions and execution rather than leaving those controls to individual local setups. Companies interested in learning more can join Retool’s waitlist, the access route it provides.