Why AI Agents Fail 70% of Tasks (And How Context Fixes It)
If you had an employee that only succeeded at 30% of the tasks you gave them, how long would you let them struggle without offering significant support? According to research from Carnegie Mellon University (CMU), that’s the success rate of even the most competitive AI agents. That figure became a 70% failure rate when cited by Asana when they announced their own built-in AI agents, AI Teammates. Common failure patterns, as outlined in the research, center on treating a task as finished before it actually is, encountering recurring obstacles when dealing with complex web interfaces, and fabricating workarounds when running into blockers.
But there’s another failure state that isn’t mentioned in the research, which contributes to the overall failure rate. In this guide, you’ll learn about this and other failure states for AI agents.
The reasons AI agents fail (according to CMU)
To get a concrete idea of the real failure rate for AI agents, the team at Carnegie Mellon University set up TheAgentCompany, an imaginary software company with a locally-hosted tool stack that included GitLab (a source-code repository), OwnCloud (an open-source alternative to Google Drive), Plane (a project management tool), and RocketChat (a chat app). AI agents were given a set of 175 tasks that were diverse and realistic enough to match real-world circumstances. Each task included an intent and a series of checkpoints the AI agent had to meet.
After running these tasks, the CMU research team found that AI agents failed 70% of tasks assigned to them, and many of these failures fit one of three categories.
Lack of social skills
AI agents in CMU’s tests sometimes struggled to fully understand tasks as described by colleagues, which led to them assuming a task was completed before they actually reached the final milestone. For example, an agent that needs additional context for a task would ask a colleague who it needs to go to for that context. When given the name of that person, the agent will consider its task done because it got a name— meaning it never actually completes the task that was originally given to it.
Browsing incompetence
The main advantage of AI agents is their ability to act independently, similarly to how a human coworker can. But whenever a task involved browsing the web or otherwise using web-based interfaces, failure rates tended to increase. Web interfaces are complex to navigate, and AI agents sometimes run into obstacles that stop them cold. Something as simple as closing a welcome pop-up is enough of a blocker for an AI agent to fail at a task.
Fabricated workarounds
When a junior employee isn’t sure what they need to do next, they’ll usually ask a more senior collaborator. AI agents in CMU’s tests often didn’t ask for clarification when next steps weren’t clear. But they wouldn’t just stop trying to complete a task. They’d create their own fabricated shortcuts to try to get to a satisfactory resolution, without the task actually being completed. For example, an agent in CMU’s test needed to reach out to a specific colleague in RocketChat. When it couldn’t find the right colleague, it renamed a different user in RocketChat to match the name of the colleague it needed and contacted that user.
The missing failure state: Isolation
The three categories named above are the ones cited in CMU’s research. But there’s another failure state the research doesn’t mention: isolation. In CMU’s tests the AI agent was given access to every tool in the team’s simulated environment. Most organizations use a tool stack that goes beyond four tools. AI agents rarely have access to every tool an organization uses, which means they’re isolated from essential context and tasks.
This is no different from how humans work in most organizations. With companies using an average of 106 tools, people frequently switch from tool to tool just to get a single task done. Tool silos keep essential data trapped, meaning there’s no way to get a complete picture without switching between tools. Humans usually accept this as an essential part of their day-to-day work. But AI agents can’t make that switch natively; they need integrations.
Foundation AI models typically have at least some built-in integrations to mitigate this. Claude, for example, has built-in connectors for GitHub, Gmail, Google Calendar, and other tools. ChatGPT has connectors for tools like Google Drive, Box, Notion, Slack, Jira, and HubSpot. But unless you do the work to install these connectors (which can range from just clicking a button to writing your own scripts), your AI agents won’t have access to these tools by default.
Combined with the failure states and patterns already defined in CMU’s research, this isolation can compound an agent’s failure rate. If agents are already known to fabricate workarounds or consider a task done before it’s actually reached the final milestone, imagine how much more likely this is when they don’t have all the information they need. Or, put another way, how much more often would they succeed if they had all that extra context?
Humans can get that extra context independently. Agents help accessing it.
The fix: Better context in your tools
AI context depends on two essential elements: software integration and read-write capability.
Software integration is the bridge that gives AI agents access to the tools in your stack. Much like CMU’s team built a stack AI agents could access to carry out their tasks, the agents you use need access to your tools. It’s not just about centralizing data in a single platform; it’s about giving your agents as much access as you can while maintaining data security. Software integration comes in many different forms, from built-in connectors to one-way automation and two-way sync. Platforms have different levels of complexity, with some being simple enough that anyone can quickly build and deploy their own integrations, while others require technical teams and coding knowledge. To properly give AI agents the right level of context, you need software integrations that actually fit your needs, both in complexity and in the amount of data they can sync.
Read-write capability is part of this. Think about it this way. When a human teammate finds outdated information in a tool, what do they do? They either update it themselves or flag it to someone else who can. If AI agents don’t have that capability, they’ll be constantly working with outdated information. That doesn’t mean agents need the ability to make updates in every tool you use, but software integrations should be able to update data in every relevant tool. Otherwise, you’ll never have up-to-date information.
Where Unito comes in
Unito is a two-way sync platform with integrations for some of the most popular tools on the market, from project management apps to software development tools and spreadsheets. Two-way integrations mean Unito can keep all your tools up-to-date as you work, eliminating the manual copying and pasting work involved in most cross-tool workflows.
So how does that keep AI agents from making mistakes?
Imagine an AI agent that’s used to hand off support tickets from your customer support tool to an IT team. By default, in most organizations, that agent might only have access to your customer support tool. That means they get all the information they need from the initial ticket and the conversations your customer support team has with customers. But they don’t have access to anything in the tool your IT team uses. That means they don’t have any context on how the IT team works, what they actually need from a handoff, or any questions they have after the handoff.
With Unito, you can sync the IT team’s tool with your customer support tool, pairing the tickets in each one. Even if your AI agent only has access to tickets on the customer support side, it still get to see questions from IT (as they become comments in the customer support tool) and insights into how they work. That gives it more context into how this work happens, leading to more autonomy and a better success rate.
Want to see how Unito can improve your AI agents?
Meet with a product expert for a demo.
FAQ: Why AI agents fail
How often do AI agents fail?
According to research from Carnegie Mellon University, AI agents can only complete 30% of tasks autonomously. That means they have a 70% failure rate without the proper context or support.
Why do AI agents fail on multi-step tasks?
Each step in a task has its own risk of failure, and these risks compound for multi-step tasks. AI agents don’t usually double-check their work between steps, meaning an early error can lead to a cascade of failures throughout the rest of a task.
Can shared context reduce AI agent failures?
Shared context doesn’t fix all potential agent failures. It can’t prevent hallucinations, fix misread file formats, or make up for unclear instructions. It can fix failures due to outdated information or missing data from a siloed tool.
Is isolation the main cause of AI agent failure?
Isolation isn’t the main cause of AI agent failure, but it can cause a significant portion of failures. When AI agents only have access to a few tools, they develop blind spots that make answering your questions or acting autonomously more challenging.