Why Purple Teaming Is the Future of Cyber Defense
Learn how purple teaming improves threat detection, response speed, and cyber resilience.
In this segment of “Office Optional with Larry English,” Larry explores the five common failure modes preventing organizations from successfully scaling AI agents and explains how intentional design, clear structure and strategic oversight can turn pilot projects into real business results.
Most organizations are stuck in the gap between experimenting with AI agents and realizing real business impact.
While adoption is accelerating, the results are lagging. A November 2025 study from McKinsey found that approximately 60% of companies have begun experimenting with AI agents, yet fewer than a quarter have scaled the technology meaningfully. What’s more, Gartner analysts estimate that more than 40% of agentic AI projects will be scrapped within the next two years.
Leaders know they need to figure out how to incorporate AI agents into their organizations, but they haven’t quite figured out how to move past the pilot phase. Yet this is a problem that can’t wait to be solved. Organizations that are making moves now are on the path to experience compounding gains in productivity, cost efficiency and decision speed. Those that drag their feet face an ever-widening gap.
Five Failure Modes Holding AI Agents Back
Failed agent initiatives are frustrating, but they’re also a strategic liability. What causes agent project failure? And what can leaders do to avoid agent project breakdown? Here are five common AI agent failure patterns and insights on how to design AI agents for success:
1. Expecting AI agents to be able to do everything.
AI agents are powerful tools—when they’re in a tightly constrained environment. That caveat is crucial. When organizations try to build agents that do too much, the result is often failure.
AI agents won’t perform well when given broad responsibilities without specific operating parameters. However, the mistake isn’t usually as obvious as telling an agent to “handle customer service.” It’s more subtle, such as telling an agent to “help with onboarding” without defining which onboarding steps, which data sources to use, which decisions require a human and which paths are off-limits.
Agents will fill every gap in instructions with their own judgement, and that judgement isn’t always what you’d expect. They require a narrowly defined scope with very specific instructions and clear boundaries.
Take invoice processing. Building an agent to “automate accounts payable” won’t work because that’s not a task, it’s an entire department. An agent given that broad mandate will try to handle vendor communications, exception routing, approval workflows, payment execution and more all at once, and it will do all this poorly.
In comparison, the organizations seeing impact are designing agents with surgical-level scope, such as a single agent that extracts line items from invoices and matches them against purchase orders. That agent does just that one task; it doesn’t chase down approvals or email vendors or decide how to handle exceptions. Agents designed with this level of specificity and limited scope work on day one and keep working on day 300.
2. Designing AI agents without considering edge cases.
In controlled environments like pilots, inputs are predictable. In the real world, they’re anything but. Edge cases multiply quickly, and performance drops off a cliff. Designing AI agents without considering edge cases is a huge issue, and it’s what keeps many agents in pilot purgatory.
The issue with edge cases compounds when you give agents memory. Agents begin to “learn” from past interactions, including attempted workarounds. It’s important to note that agents will try everything in their power to accomplish the task they’ve been given, including less-than-ideal workarounds. Over time, these exceptions to the rule can become the agents’ default behavior.
Instead, design agents for failure, not just success. It’s impossible to fully anticipate every situation an agent will encounter, so you need to design explicit off-ramps. Before deployment, conduct “edge case” exercises, with the goal of generating AI agent guardrails and instructions for telling agents what not to do, what paths are off-limits and what scenarios require escalation to a human.
3. Failing to align problem structure and agent architecture.
Every problem has a structure. Problems can be linear, iterative, exploratory or conditional, to name just a few examples. Yet many organizations make the mistake of taking a one-size-fits-all approach with AI agent design.
That’s like using the wrong tool for the job, grabbing a screwdriver when what you really need is a hammer. It might work occasionally, but it will be inefficient and frustrating.
Instead, take care to align agent architecture with the type of problem they’re solving. If the problem is iterative, your agent should be too. If the problem is audit-driven, design for validation and verification.
For instance, app modernization efforts often resemble an archeology expedition. You’re constantly uncovering new information that causes you to reinterpret and adjust course. This kind of problem requires an adaptive agent design, not a rigid, linear workflow.
4. Lacking stewardship over AI agent costs.
Poorly designed AI agents can get stuck in loops, repeatedly processing without producing valuable outputs. Additionally, agents delivering similar outcomes can vary in cost by multiples—it all depends on how they’re processing the requests.
With token-based pricing, this means that organizations can accidentally waste a lot of funds without getting any results.
To avoid this risk, establish AI agent cost stewardship at the individual level by implementing clear instructions and governance around AI agent use. Employees need a baseline of AI agent literacy. They need to understand how agents operate and the cost implications of using AI agents. They also need clear stopping conditions and guidelines on monitoring for looping or redundant behavior.
5. Expecting AI agents to deal with information chaos.
With AI agents, more data doesn’t mean better outcomes. The effectiveness of an agent is directly tied to the clarity of its inputs. Yet many organizations throw everything they have at agents and expect them to figure it out. This doesn’t work, because too much unstructured information just leads to noise.
Instead, be intentional about context. Curate and structure the information agents receive to ensure a straightforward path for agents to take.
For example, a sales AI agent fed an entire CRM history will struggle. One given a structured brief that includes customer segment, last interaction, and recommended next action will perform more reliably.
Why AI Agent Performance Is A Communication Issue
One of the most overlooked aspects of AI agent design is communication. Organizations tend to focus on the technical aspects of models and architecture, but they neglect thoughtful communication design.
It’s critical to understand that large language models don’t just process data, they learn patterns of human communication. That makes how you instruct, guide, and constrain them just as important as the underlying technology. This is why agent design is as much an operating model discipline as a technical one
For example, constant overpraise can reinforce overconfidence in an AI agent, leading the agent to resist correction. Likewise, too much criticism can have the opposite effect, causing hesitation or deferral. In both cases, the AI agent is adapting to the behavioral patterns you’ve established.
AI agents aren’t failing because the technology isn’t ready. They’re failing because organizations are treating them like tools instead of entire systems that need structure and intentional design.
To move beyond the pilot phase, companies must design AI agents with narrow scope, align AI agent architecture to the problem, manage costs with individual-level stewardship and treat communication as a core part of agent design.
