Many enterprises face challenges when increasing AI agent utilization. In fact, one McKinsey survey found that 20% of respondents reported AI-related operating costs, such as token usage, as constraints on AI implementation. Yet despite these impediments, 40% of the large organizations surveyed are actively scaling up this technology.
As enterprises continue to invest in agentic AI, one path to improving AI ROI lies in identifying ways to mitigate these limiting factors. While many begin by comparing agent models to reach this goal, recent industry discussions have focused on a different architectural consideration: the agent harness.
CIO and The Wall Street Journal suggest that this software layer, which orchestrates AI agent actions, can directly impact agent ROI. With large-scale AI deployments underway, enterprises should understand how exactly these harnesses influence outcomes and determine the harness capabilities they require to achieve measurable value from their technology investments.
The Harness Effect Is Measurable
Several preprints and benchmarks have begun exploring the effects of the agent harness on AI ROI. An evaluation of these studies can help determine certain factors that a harness influences.
For example, a study centered in part on costs and harness design found a five- to thirtyfold difference in cost per successful task, even when researchers held the task, prompt, and model constant. This means the harness plays an important role in token and tool usage and is therefore a direct lever for cost-effectiveness.
A benchmark evaluating a Claw-SWE agent harness across various coding tasks demonstrated a related effect. This research found that harness choice could influence task performance by 27.4 percentage points. This suggests that the harness layer can shape how effectively an agent carries out a workflow, even when the underlying model remains fixed.
Risk is another factor that appears dependent, in part, on harness selection. An experiment found that altering the harness-level authorization architecture reduced execution of injected effects from 33–47/75 runs to 3/75. That means exposure extends beyond the model itself.
The findings indicate that harness design can not only influence cost-effectiveness and performance but also an organization’s ability to control long-term risks.
Yet despite the importance of harness-level controls, McKinsey found that only 30% of organizations have reached a maturity level of three or higher across the dimensions of strategy, governance, and agentic AI controls. Closing this maturity gap will involve configuring harnesses around several capabilities that affect cost, behavior, and risk.
Set the Rules Around Agent Action
To begin controlling agent performance, organizations should consider certain baseline harness configurations that address:
- Cost:T
For organizations to mitigate unnecessary agent costs, organizationstheyshould embed boundaries around tokens, tools, prompts, and calls within the harness layer. This involves defining appropriate limits and thresholds for each category based on the task’s requirements. - Behavior: To obtain effective AI outputs, organizations require parameters for agent behavior within the harness. This means defining the sequence of steps an agent can follow, when it can use a particular tool, how it verifies outputs, and when it escalates a given task.
Both cost and behavior play into organizational risk, meaning that setting boundaries around them can help limit exposure. However, there are certain risk-specific harness parameters an organization may benefit from.
Build Controls Around the Stakes
Classifying each task by risk level can help mitigate exposure. Organizations can identify which actions carry the highest risk and establish more stringent controls based on an agent’s level of autonomy and access to enterprise systems. For example, customer-facing outputs, HR or financial decision mechanisms, regulated workflows, or sensitive data processing will likely require more intensive guardrails than internal research or note summarization.
For higher-risk tasks, governance, risk, and compliance (GRC) professionals can configure the harness to restrict system access and incorporate additional legal and compliance reviews. In contrast, lower-risk actions may require fewer review points and can often operate under basic monitoring and validation requirements.
Alongside risk classifications, harnesses should also document actions, exceptions, approvals, decisions, and incidents. That is how enterprises determine ownership and maintain evidence for audit, internal review, and leadership reporting.
By building parameters around cost, behavior, and risk, organizations can enable agents to execute workflows while limiting unnecessary expenses and the exposure of sensitive information exposure. These harness guardrails essentially give GRC professionals greater insight into agent processes and help them strengthen ROI.
Make Visibility Part of AI Scale
More than 60% of organizations expect to deploy AI agents in the next two years, according to Gartner. Making these initiatives successful and extending them beyond the pilot phase will likely depend on how well an organization can control costs, behavior, and risk.
Agent harnesses are one useful lever of this strategy. They allow organizations to scale their visibility into how agents reach and execute actions, rather than limiting their insights to only the final outcome. By configuring this architecture to incorporate meaningful controls, enterprises can better align agent actions with measurable ROI.
