Enterprise AGI Was Achieved in February 2026 | Bernardo Aceituno

Summary · 2 min read

Aceituno cofounded Stack AI and serves as its president; the company helps enterprises build custom AI-agent systems for internal back-office processes. The hosts say Asana acquired Stack AI earlier in 2026.

Aceituno says enterprise AGI arrived in February 2026 for a defined set of automation tasks, where agents reached human-quality reliability across five capabilities. He says the challenge has shifted from prompting and APIs to governing access, deployments and agent actions. He argues that open-weight models give enterprises more control over cost, hosting and model versions, while websites built for people still obstruct agents. He expects agents to appear across more workplace software and become more proactive.

Minute by minute

1. 1:02

   Enterprise automation reaches a threshold. Aceituno says Stack AI helps enterprises build custom agents for back-office work, including in regulated and on-premise environments. He says five capabilities—understanding and querying unstructured data, orchestrating tools and actions, using visual interfaces, and synthesizing information—reached human-quality reliability by early 2026.

2. 3:04

   AGI depends on the task definition. Aceituno defines enterprise AGI as an agent reliably completing a specified task at human quality when given the right problem and tools. He says this does not mean personal or video AGI has arrived, and that there is still room to improve in those areas.

3. 4:00

   Open weights reduce provider lock-in. Aceituno says enterprises feel less pressure to partner deeply with frontier labs as open-weight models improve. He argues that businesses can consider alternatives to the providers tied to their cloud platforms, while still using the latest hosted models when they choose.

4. 5:54

   Human-oriented websites block agents. Aceituno says websites often block agents with bot checks and CAPTCHAs, making it hard for them to scan pages or complete tasks such as shopping. He expects marketplaces and social networks to become more agent-friendly, but says they are not there yet.

5. 7:01

   Agent-first services need new interfaces. Aceituno says models have become good at using visual interfaces, so the remaining challenge is adapting services for agents. He imagines an agent-first forum using a simple MCP server with clear commands and navigation, plus ways to verify an agent's quality, intent and operator.

6. 8:27

   Legacy systems can be bridged. Aceituno says banks can connect agents to IBM mainframe KYC systems through a tunnel and scripted computer use, even when those systems remain unchanged. He says external commerce and communication tasks are harder, while tools such as Asana are becoming places where workers and agents share a system of record.

7. 10:38

   Enterprise AI needs central governance. Aceituno says teams can build separate AI applications without clear ownership of hosting, access or credentials. He argues that CIOs need to control who can create, deploy, manage and access those tools, and to scale them or shut them down; he presents Stack AI's collaborative no-code platform as one way to do this.

8. 14:27

   Open weights passed an enterprise threshold. Aceituno says open-weight adoption surged after models became capable enough for production tasks at a manageable size. He cites Kimi K3 as a model that drew attention beyond AI practitioners, and says cost, hosting, version control and control over upgrades all matter to enterprises.

9. 17:30

   Agents will become more proactive. Aceituno expects agents to become available across work tools such as Teams, Asana and Confluence, and to monitor operations and suggest or take actions. He imagines agents flagging 300 spam emails or identifying a promising job candidate without waiting for a prompt.

Published Oct 17, 2026.