AI firms are pushing agents as the following major disruptor on the workplace, but experts say they’re still Not ready for the essential time. AI agents often have difficulty making decisions themselves, hallucinating continuously, cannot work with other agents. fail within the event of confidentiality awarenessand integrate bad in existing systems.
Industry pioneers like Andrej Karpathy And Ali Ghodsi said that like using autonomous vehicles, just like the use People should be within the loop So that agents may be successful.
A startup called mixture I would really like to deal with this with the AI agent platform, which not only keeps people within the workflow, but may also interact with agents directly from their e -mail or paint.
“We meet customers where they’re today,” Elliot Katz, co -founder of Mixus, told Techcrunch. “Where is one and all within the workforce today? For probably the most part they’re in e -mails. And in order that we will do that by e -mail, we consider that we will democratize access (on agents).”
If Mixus works reliably, this may solve an enormous problem within the AI agent room. Most AI firms today provide you with either a prefabricated assistant, à la Chatgpt or Gemini, or developers should construct customer-specific agents with framework conditions comparable to Langchain, Autogen or Crewai.
Mixus was only began in Beta from Stanford at the tip of 2024, but has already collected $ 2.6 million in pre-resettling financial resources and brought some customers, including rainbow stores of the outfitters, in addition to others about finance and technology.
The startup says that its biggest sales argument is user -friendliness, from the best way you possibly can create agents, to the best way you possibly can interact with you. Users can use text instructions to establish their agents on the Mixus platform via a chat function or just by e -mail to instructions to agent@mixus.com. Then created, drives and manages single or multi-step agents directly from the inbox.
Techcrunch event
San Francisco
|
twenty seventh to October 29, 2025
For example, a customer can use support managers to make use of a command prompt that reads like this:
Create an agent who finds all open tasks in Jira in Project Mixus dummy and send me a report with information on all overdue tasks. Design e -mails to all agents which have overdue tasks and have me checked in chat and with easy clear formatting for e -mails (no attachments/documents). As soon as I even have checked, send the e -mails. Do it now. And go forward, run it every Monday at 7 a.m. PST.
Katz and his co -founder Shai Magimof brought the agents for Techcrunch and showed add human verifiers for his or her agents by simply instructing what step the agent should ask for supervision.
For example, they carried out an agent to research Techcrunch reporters before they set them up. The agent identified and picked up technologies and trends, analyzed the data to discover potential story angles, and created a research report by which the outcomes were summarized. In the last stage, the agent was instructed to send the data to Katz for checking. After approval, the agent would send the finished research report back to Magimof.
The founders found that folks within the loop may be as much or as little as necessary-Magzimof that organizations can arrange company-wide rules, e.g. B. to make sure that an email is checked by an individual if it is distributed outside the corporate.
The introduction of other colleagues into the workflow is so simple as marking in chat with an AI agent or copying e -mails to the agent. This is one other outstanding place in comparison with agents within the markets today: Most models are single users, and while the AI and Slack enable users to work together in common rooms, they don’t let the AI conversations and tasks between teammates manage in real time.
Another core function of Mixus is the flexibility to commemorate files, chats, input requests and agents.
“We have created rooms in order that every team, one and all, every group of individuals can have a typical memory,” said Magimof. “Then all my agents, all my files, may be all people on this memory of this memory.”
While Chatgpt and Claude each support the memory, their corporate plans don’t yet support the common agent storage for users.
What else can Mixus do?
In our interview, the founders carried out a one -hour demo that showed quite a few applications and skills. Mixus' agents appear to be able to, which reflects a high degree of autonomy and memory that the corporate sets at the highest of the AI agent spectrum. This signifies that the product works as reliably as within the demo.
Like other agents, Mixus can integrate into other tools, from Mail Mail to Jira, and users could make agents perform immediately or in response to a schedule. Agents can perform and edit documents or spreadsheets – much like Chatgpt, Microsoft Copilot and Google Gemini. However, these are sometimes limited to sandbox environments.
With Mixus, agents may also autonomously navigate the organizational context. B. to search out out who has a task in a company by seeing Jira tickets through.
Mixus agents are based on a mixture of Claude 4 by Anthropic 4 and Openaai. You even have access to the online, which Magzimof may be shown for tasks comparable to live research or monitoring. He described it as “Google warnings about steroids”.
Together, Mixus appears to be a less productivity tool than more like a tireless digital colleague – one other ambitious try to reinterpret the AI as an worker. If it really works as announced, your next “worker” might not be human, but your inbox could get faster than you.
