CraftOS Gives Its Own Agent More Work, With Tighter Limits

The most persuasive evidence that an AI agent works inside a company is not a demo video but how much real work it can carry. An interview with KoreaTechDesk offers one internal datapoint from CraftOS, a Japan-based AI startup.

CraftBot, the company's agent, has gradually taken on more administrative and operational responsibilities. Co-founder and CEO Dr. Tham Yik Foong said that around three months into development the agent could handle simple tasks but performance was inconsistent: team members often had to supervise its work, restart processes or review results extensively before relying on them.

That limited how much responsibility the company could practically delegate, Tham said. As CraftBot matured, his informal assessment of internal workflows put the share of operational tasks assigned to the agent at roughly 60%, up from about 30%.

Several things moved together to produce that shift: CraftBot itself became more stable, the underlying language models grew more capable, and CraftOS built an internal Company OS on its own Agent App framework to give the agent a structured environment for company information and everyday workflows.

The Company OS pulls operational records, task information and business processes into an application CraftBot can interact with directly. In Tham's words, it gives the agent a more deterministic environment for company data and workflows, cutting the chance of misreading an instruction or losing track of a task and reducing the need for repeated human intervention.

That points to an engineering fact that is easy to overlook: better model reasoning does not by itself produce dependable business automation. When information is scattered across spreadsheets, conversations and separate applications, every additional interpretation is another chance for confusion or a human rescue. Defining records, task statuses and available operations narrows that space.

CraftBot currently handles lead sourcing, customer relationship management, event organization and document preparation. Those responsibilities carry different levels of independence, since organizing information or drafting documents produces output without immediately changing the company's relationships or commitments, while software development remains largely human. The CraftOS lesson is less about granting autonomy than about buying more delegable work with clearer boundaries.