Decawork

Decawork is a control layer for AI agents that decides, in real time, whether each action an agent attempts should be allowed. Rather than relying on static permissions, it trains a small model specific to each company that approves or blocks actions based on the company's policies, the task the agent is working on and the actions it has already taken. It is built for organizations running agents across many teams, from marketing and recruiting to IT, finance, legal and engineering, that need oversight of what those agents do. Agents are connected by routing their API requests and MCP tool calls through Decawork, which keeps an agent registry with each agent's identity, owner, tools and activity in one place. Policies are written in plain English and define both what the model should enforce and when a person must step in. Access can be granted for the task at hand, reassessed as the agent works and revoked when the task is done, so the same action can be approved in one context and refused in another. The platform works with agents built on any stack, and the company states it holds SOC 2. Pricing is not published. The site offers a demo booking and does not list plans or amounts, so access appears to be sales led. Decawork states that it is backed by Y Combinator.