AI companies push representatives as the next big job harasser, but experts say they are still Not ready for the main time. AI -agents often struggle to make decision -making, often as hallucinates, cannot collaborate with other representatives, fail in confidentiality consciousnessand integrate poorly into existing systems.
Industrial pioneers such as Andrej Karpathy and Ali Ghodsi has said that such as the introduction of autonomous vehicles, People have to be in a loop To the success of agents.
Startup called Blend Wants to intervene in its ai -agent platform, which not only keeps people in workflow, but also allow users to interact with agents directly from email or loose.
“We meet customers where they are today,” Elliot Katz, founder of Mixus, told Techcrunch. “Where is each person in the workforce today? For the most part, they are via email. And because we can do this by email, we believe that it is the way we can democratize access (representatives).”
If Mixus works reliably, it can solve a big problem in AI agent mode. Most AI companies either give you a pre-built assistant à la chatgpt or twins or developers must build customized representatives using frames such as Langchain, Autogen or Crewi.
Mixus was not launched in the beta area from Stanford until the end of 2024, but has already raised $ 2.6 million in funding and brought some customers, including the Rainbow stores in the clothing store chain, as well as other funding and technology.
Startup says its biggest outlet is the ease of use of how it helps you create representatives of how you can interact with them. Users can use text objects to determine their representative on the Mixus platform via the chat function or simply by sending an email to agent@mixus.com. Then Mixus builds, drives and controls one or multi-stage agents directly from the mailbox.
TechCrunch event
San Francisco
And
27.-29. October 2025
For example, the seller may use a prompt that reads so:
Create an agent that will find all the open tasks of JiRa’s Project Mixus and send me a report of any late task. Nature to email all tasks with late tasks and ask me to check them in chat and simple clear email (no attachments/documents). When I check, send emails. Run it now. And move on, drive it every Monday at 7am pst.
Katz and his founder, Shai Magzimof, are demonstrating Techcrunch representatives, showing how to add to your representatives the human verifiers simply by teaching the agent to ask you to control you.
For example, they kept a representative to do research on Techcrunch journalists before they rise. The agent identified and compiled technology news and trends, analyzed information to identify possible story angles, and prepared a research report that summarizes the findings. In the final step, the agent was advised to send information to Katz for certification. Once approved, the agent will send a completed research report to Magzimof.
The founders pointed out that people may be in the loop as much or as little as needed-Magzimof said that organizations can set company-wide rules, such as ensuring that one checks the email if it is shipped outside the company.
Bringing other colleagues to the workflow is as easy as it means to chat with an AI agent or copy them via email to the agent. It is another standout compared to market representatives today: most models are one of the users, and while AI and Slack give users the opportunity to cooperate in shared spaces, they do not allow AI to control conversations and tasks between teammates in real time.
Another Mixus core feature is its ability to remember files, discussions, prompts and representatives.
“We created the premises so that every team, every person, each group of people may have a common memory,” Magzimof said. “Then all the representatives, all my files, all people can be in the memory of that particular state.”
Although chatgpt and Claude both support memory, their business plans do not yet support the shared agent memory between users.
What else can Mixus do?
In the interview, the founders ran an hour -long presentation, showing different uses and abilities. Mixus’ representatives appear to be capable, reflecting high autonomy and memory that placing a company towards the top of the ai -agent spectrum. Or if The product operates both reliably and at Demo.
Like other representatives, Mixus can integrate with other tools, from Gmail to Jira, and users can trigger agents to drive immediately or on schedule. Agents can drive and modify documents or spreadsheets – similar to chatgpt, Microsoft Copilot and Google Gemini, but are often limited to sandboxes.
Mixus also allows agents to navigate independently with the organization – such as finding out who owns the task in the organization by finding Jira tickets.
Mixus agents have also been given to the network, which, according to Magzimof, is built by the combination of Antrus agents from the combination of Antropic Claude 4 and Openai’s O3, which Magzimof, according to Magzimof, can be utilized, such as Live study or monitoring. He described it as “Google Alerts in Steroids”.
Together, Mixus seems to be a less productivity tool and more than a tireless digital colleague – another ambitious attempt to describe AI as a partner. If it works advertised, your next “colleague” may not be a person, but it may get through the mailbox faster than you.
Correction: This article was updated to reflect how much Mixus has been raised so far.
Do you have a sensitive tip or confidential documents? We report the internal activities of the AI industry – companies that form their future to the people that their decisions affect. Reaches Rebecca Bella Rebecca.bellan@techcrunch.com and Maxwell Zeff maxwell.zeff@techcrunch.com. For safe communication you can contact us via signal @Rebeccabellan.491 and @Mzeff.88.
