AI Agents & Email: A New Approach by This Startup

The Current Limitations of AI Agents

Many AI firms are promoting agents as a transformative force in the workplace, yet industry analysts suggest they aren't currently prepared for widespread adoption. A common challenge with AI agents is their difficulty in independent decision-making. They frequently exhibit a tendency to hallucinate information, struggle with collaborative interactions, demonstrate insufficient confidentiality safeguards, and present integration issues with pre-existing systems.

The Need for Human Oversight

Leading figures in the field, including Andrej Karpathy and Ali Ghodsi, have emphasized the necessity of human involvement for successful agent deployment. This approach mirrors the current state of autonomous vehicle technology, where human supervision remains crucial.

Mixus: An Agent Platform Focused on Accessibility

A new company, Mixus, is aiming to address these challenges with an AI agent platform designed to integrate humans directly into the workflow. Furthermore, Mixus allows users to interact with agents through familiar channels like email and Slack.

Democratizing Access to AI Agents

“Our strategy is to engage with customers in their current work environment,” explained Elliot Katz, co-founder of Mixus, in an interview with TechCrunch. “The majority of the workforce relies on email, and by leveraging this existing infrastructure, we aim to broaden access to AI agents.”

Solving a Key Problem in the AI Agent Landscape

If Mixus proves to be dependable, it could resolve a significant issue within the AI agent sector. Currently, most AI solutions offer either pre-configured assistants, such as ChatGPT or Gemini, or require developers to construct custom agents using frameworks like LangChain, AutoGen, or CrewAI.

Early Traction and Funding

Launched in beta from Stanford in late 2024, Mixus has already secured $2.6 million in pre-seed funding. The company has also onboarded initial customers, including the clothing retailer Rainbow Shops, and organizations within the finance and technology industries.

Ease of Use as a Core Advantage

The startup highlights the simplicity of its platform, both in agent creation and interaction, as a primary benefit. Users can define agents using natural language prompts within Mixus’ interface, or simply by sending instructions to agent@mixus.com. Mixus then handles the building, execution, and management of both single-step and multi-step agents directly from the user’s inbox.

Example Use Case: Customer Support

Consider a scenario where a customer support manager utilizes a prompt to define an agent’s task:

Human Verification and Agent Workflow

Katz and co-founder Shai Magzimof demonstrated the platform to TechCrunch, showcasing the ability to incorporate human reviewers into the agent workflow. This is achieved by instructing the agent to request oversight at specific stages of a process.

For instance, an agent was used to research TechCrunch reporters prior to outreach. The agent gathered relevant news and trends, analyzed the data for potential story ideas, and compiled a research report. At the final step, the agent was instructed to forward the information to Katz for approval. Upon verification, the report was then sent to Magzimof.

Flexible Levels of Human Involvement

The founders emphasized the adaptability of human involvement, noting that organizations can establish company-wide rules, such as requiring human review for any email sent externally.

Collaboration and Real-Time Task Management

Integrating colleagues into the workflow is straightforward, involving simply tagging them in agent chats or copying them on emails to the agent. This feature distinguishes Mixus from many existing agents, which are typically designed for single-user operation. While platforms like Notion AI and Slack offer collaborative spaces, they don’t enable AI-driven conversation and task management between team members in real time.

Persistent Memory for Enhanced Agent Performance

A key capability of Mixus is its ability to retain information about files, chats, prompts, and agents.

“We’ve developed Spaces to provide a shared memory for teams, individuals, and groups,” Magzimof explained. “This ensures that all agents, files, and users operate within a consistent memory context.”

Shared Memory Compared to Competitors

Although ChatGPT and Claude offer memory features, their enterprise plans currently lack support for shared agent memory across multiple users.

Exploring the Capabilities of Mixus

During a comprehensive interview, the company’s founders presented a detailed, hour-long demonstration showcasing the diverse applications and functionalities of Mixus. The agents developed by Mixus demonstrate a notable level of independence and retention, positioning the company among the leading innovators in the AI agent field.

However, this assessment is contingent on the product consistently performing with the same reliability observed during the demonstration.

Similar to other AI agents, Mixus facilitates integration with a variety of existing tools, including Gmail and Jira. Users have the flexibility to initiate agent tasks either immediately or on a predetermined schedule.

Agents are capable of directly operating on and modifying documents and spreadsheets – a feature comparable to ChatGPT, Microsoft Copilot, and Google Gemini, but often without the limitations of isolated testing environments.

Autonomous Navigation of Organizational Data

A key feature of Mixus is its ability for agents to independently interpret organizational structures. For example, agents can determine task ownership within an organization by analyzing Jira ticket information.

The agents leverage a combined architecture utilizing both Anthropic’s Claude 4 and OpenAI’s o3 models. Furthermore, Mixus agents possess web access, which, according to Magzimof, enables tasks such as real-time research and monitoring.

He characterized this functionality as a significantly enhanced version of Google Alerts.

A New Paradigm for AI Collaboration

Collectively, Mixus appears to transcend the boundaries of a typical productivity application. It presents itself as a persistent digital assistant – representing another ambitious effort to redefine AI as a collaborative partner.

Should it function as promised, future colleagues may not be human, but could potentially manage your inbox with greater efficiency.

Note: This article has been updated to accurately reflect the total funding raised by Mixus.

Do you have confidential information or a sensitive tip to share? We are dedicated to reporting on the AI industry, covering both the companies driving its evolution and the individuals affected by their choices. Please contact Rebecca Bellan at rebecca.bellan@techcrunch.com and Maxwell Zeff at maxwell.zeff@techcrunch.com. For secure communication, reach out via Signal to @rebeccabellan.491 and @mzeff.88.