A New Gatekeeper Emerges for Corporate AI Agents
Decawork, a startup from Y Combinator's Summer 2026 batch, has launched a control plane designed to give enterprise IT departments centralized oversight of the AI agents proliferating across their organizations.
The San Francisco company joins a suddenly crowded field. Between March and June, IBM, Salesforce, OpenHands, Galileo, and Atomicwork all shipped similar products aimed at solving what has become an acute problem for corporate technology teams: how to manage autonomous software that can read Slack channels, update Salesforce records, and close Jira tickets without human supervision. Decawork secured SOC 2 certification on July 20, a credential that matters when selling into enterprise security committees.
The underlying challenge is straightforward, if unsettling. Companies have spent two decades building identity and access systems for employees. Then AI agents arrived, and many organizations reverted to a cruder approach: sharing API keys. Sarthak Aggarwal, Decawork's founder, put it bluntly in a LinkedIn post: enterprises went "back to pasting API keys."
Decawork's product sits between these agents and the systems they touch. Rather than letting an agent authenticate with a shared credential, the platform assigns each one its own identity and scoped permissions. When an agent tries to execute a tool call, the request routes through Decawork's gateway. Sensitive actions pause for human approval. Everything else gets logged with full attribution. "Agents never hold raw credentials," the company says on its security page. "Nothing runs until IT approves."
The workflow unfolds in three stages, according to Decawork's documentation. During deployment, each agent receives an identity and defined access permissions. Launch is gated behind a checklist covering monitoring, spending caps, and IT sign-off. Once live, the governance layer routes all tool calls through the control plane, holding certain actions for manual approval while logging the rest. The maintenance phase watches for credential expirations or integration failures. In one scenario the company describes, an agent pauses automatically when credentials expire and IT receives a Slack alert in a designated channel.
Teams can connect agents they've already built on platforms like Claude Code or Codex. Decawork provides the operating layer without forcing a rebuild.
The market timing reflects a broader pattern. As agent frameworks matured over the past year, control planes emerged as the next infrastructure layer. IBM introduced its Agentic Control Plane within watsonx Orchestrate in June. Salesforce expanded Agent Fabric with what it termed a "trusted agent control plane" in May, a few months after the initial Agent Fabric debut. OpenHands and Galileo launched their versions in early spring, with Galileo opting for an open-source model. Atomicwork announced a "workforce control plane" for what it calls AI coworkers. LangChain introduced its LLM Gateway with similar concepts in mid-May.
By July, Onepane had published an explainer defining the category itself, a signal that the market had reached a certain threshold of activity. The post outlined core concepts: identity-bound execution, runtime policy enforcement, observability.
Aggarwal has leaned into the security argument. In one post, he cited figures claiming that 88 percent of companies experienced an AI agent security incident in the past year, and only 7 percent have designated someone accountable for agent oversight. The post did not identify the source of those statistics, though the general anxiety among IT leaders is real enough. Agent security incidents remain underreported, and accountability structures are often unclear.
Decawork's SOC 2 examination, completed on July 20 by Securance Pro Assurance PLLC, covers security, availability, and confidentiality. The report is available to customers and prospects under confidentiality agreements, a standard arrangement in enterprise sales.
The company emerged from Y Combinator's most recent batch. Aggarwal acknowledged investors Brad and Christina in a LinkedIn post, though he did not provide their full names. LinkedIn indicates the company employs between two and ten people, including Aggarwal, Aman Raj, and Vipul Tomar.
Whether the control plane category consolidates or fragments depends partly on how enterprises choose to deploy agents at scale. Some may prefer integrated suites from IBM or Salesforce. Others might opt for standalone platforms like Decawork that promise flexibility across different agent frameworks. For now, the rush of product launches suggests that vendors believe the problem is urgent, and that IT departments are looking for answers.
