Join the event that trusts business leaders for almost two decades. VB Transform brings together people who build a real business AI strategy. Learn more
While companies are faced with the challenges of the deployment of AI agents in critical applications, a new more pragmatic model emerges which makes humans in control as a strategic backup against AI failure.
Such an example is MixusA platform that uses a “colleague in a loop” approach to make agents of IA reliable for critical mission work.
This approach is a response to growing evidence that fully autonomous agents are a bet with high issues.
The high cost of uncontrolled AI
The problem of Hallucinations Ai has become a tangible risk because companies explore AI applications. In a recent incident, the cursor of the code editor powered by AI saw his own support bot invent a false policy Restricping subscriptions, triggering a wave of public cancellations from the customer.
Likewise, Fintech Klarna company reverse course When replacing customer service agents by AI after admitting the move, led to lower quality. In a more alarming case, the commercial chatbot powered by New York AI advised entrepreneurs of engage in illegal practiceshighlighting the risk of catastrophic compliance of non -monitored agents.
These incidents are symptoms of a greater gap. According to a Salesforce from May 2025 search documentThe main agents of today only succeed 58% of the time on tasks in a single step and only 35% of time in several stages, highlighting “a significant difference between the capacities of current LLM and the requests for multiple facets of corporate scenarios of the real world”.
The loop colleague model
To fill this gap, a new approach focuses on structured human surveillance. “An AI agent should act in your direction and on your behalf,” Elliot Katz, co-founder of Mixus, told Venturebeat. “But without integrated organizational monitoring, fully autonomous agents often create more problems than they solve.”
This philosophy underpins the Mixus in a loop colleague model, which incorporates human verification directly into automated workflows. For example, a large retailer could receive weekly reports from thousands of stores containing critical operational data (for example, sales volumes, working hours, productivity ratios, remuneration requests). Human analysts must spend hours manually examining the data and making decisions according to heuristic. With Mixus, the agent of AI automates heavy lifting, analysis of complex models and signaling anomalies such as unusually high salary requests or aberrant productivity values.
For decisions with high issues such as payment authorizations or violations of policies – the workflows defined by a human user as “at high risk” – the agent stops and requires human approval before proceeding. The division of labor between AI and humans has been integrated into the process of creating agents.
“This approach means that humans are only involved when their expertise really adds value – generally the 5 to 10% critical that could have a significant impact – while the remaining 90 to 95% of routine tasks unfold,” said Katz. “You get the speed of full automation for standard operations, but human surveillance is triggered precisely when the context, judgment and responsibility count the most.”
In a demo that the Mixus team has shown for Venturebeat, the creation of an agent is an intuitive process that can be carried out with ordinary text instructions. To build an agent of verification of facts for journalists, for example, the co-founder Shai Magzimof simply described the process in several stages in natural language and asked the platform to integrate stages of human verification with specific thresholds, for example when a complaint is at high risk and can cause reputation damage or legal consequences.
One of the main forces of the platform is its integrations with tools such as Google Drive, Email and Slack, allowing the company’s users to bring their own data sources in workflows and interact with agents directly from their choice of communication platform, without having to transform contexts or learn a new interface of the publisher).
The platform’s integration capacities extend more to meet the specific needs of the company. Mixus supports the context protocol of the model (MCP), which allows companies to connect agents to their tailor -made tools, avoiding the need to reinvent the wheel for existing internal systems. Combined with integrations for other corporate software like Jira and Salesforce, this allows agents to perform complex and multiplated tasks, such as the verification of open engineering tickets and the statement of the State to a manager on Slack.
Human surveillance as a strategic multiplier
The company’s space of the company is currently undergoing a verification of reality while businesses go from experimentation to production. Consensus among many industry leaders is that humans in the loop are a practical necessity for agents to occur reliably.
The collaborative Mixus model changes the economy of AI scaling. Mixt predicts that in 2030, the deployment of agents can increase by 1000x and that each human supervisor will become 50x more efficient as AI agents become more reliable. But the total need for human surveillance will always increase.
“Each human supervisor manages exponentially more work on AI over time, but you still need more total surveillance as the deployment of AI explodes in your organization,” Katz said.
For business leaders, this means that human skills will evolve rather than disappear. Instead of being replaced by AI, experts will be promoted in roles where they orchestrate the fleets of AI agents and will manage decisions with high issues reported for their exam.
In this context, the construction of a strong human surveillance function becomes a competitive advantage, allowing companies to deploy AI in a more aggressive and safe way than their competitors.
“Companies that master this multiplication will dominate their industries, while those who hunt complete automation will find it difficult with reliability, compliance and confidence,” said Katz.
