Autonomous AI agents are moving faster than the frameworks meant to contain them. Enterprises are deploying agents that can call systems, pull data, and increasingly interact with other agents to complete multi-step tasks.
Yet few organizations can say with confidence exactly how many agents are running in their environment, what each one is authorized to touch, or who exactly is accountable when something goes wrong.
That gap can turn agentic workflows from a promising technological advancement into a potential minefield of risk unless enterprises take a new approach to governance – one that provides greater oversight into how agents are operating and what systems they can access, trust, and use.
Latest Videos FromTechRadar
Don’t wait for perfection
The market has responded to the potential threat around AI agents going rogue with a wave of new tooling: agent discovery platforms that scan for active agents, and agent harnesses that box them into approved boundaries. Both are useful, but neither solves the problem alone.
The challenge of “getting a handle” on AI agents is compounded by the pace of change. New agent tools, new AI model releases, and new orchestration options are arriving at a rate that makes any static governance model obsolete within months.
Rather than waiting for a “perfect” governance framework to emerge as a standard, organizations should take steps now to create a working structure that can evolve over time.
So, what might this look like in practice?
Continuous visibility and granular guardrails
Policies and procedures alone cannot confirm what is actually running in production. Organizations need a registration and discovery process that captures every agent in use, not just the ones teams report having built.
Real visibility comes from instrumenting the environment itself, using logging and observability data generated by the underlying models, and building analysis on top of it. Only that raw data can show what an agent is actually doing, how often, and at what cost, rather than relying on what people think it is doing.
Another key step is to move from broad guardrails to granular ones. A single, broad-based harness applied uniformly across an organization, for example, limits scope without addressing risk in a meaningful way. Different use cases call for different levels of restriction, and treating every agent identically either over-constrains valuable work or under-constrains risky work.
Guardrails need to be defined at the level of the individual agent and its specific task, not the organization as a whole. As agents begin accessing other agents to complete a task, that specificity becomes even more important: each agent needs an explicit, narrow definition of what it can do and what it can reach, so that any leakage beyond that defined scope is immediately visible.
With all of the above, enterprises should embrace an approach of “validate, don’t assume”. Putting a conceptual harness or policy in place is only the first step. Confirming that it actually holds under real conditions is a separate and ongoing exercise.
A useful approach here is to pair conceptual guardrails with a logical or physical enforcement layer: a tool that can confirm what is executable, what an agent can interact with, and whether it stays within that boundary in practice.
Continuous monitoring against policy, not a one-time sign-off, is what proves a harness is working, because agent capability and the surrounding set of tools change on a near-daily basis.
Access is not the only variable
Governance conversations tend to focus almost exclusively on what an agent can access. Equally important is the order in which it accesses information. An agent that gathers information out of sequence, or acts on data before a dependent step has completed, can produce results that are wrong even if every individual action was technically permitted.
Some workflows require steps to run in parallel; others require specific sequencing to ensure the most accurate results. Building a discrete, well-defined operational sequence for each use case, rather than allowing a model to determine its own order of operations, substantially reduces this type of error.
This is as much a process design issue as a technical one, and it relies on people defining the use case clearly before any tooling is applied.
A disciplined approach reins in agentic risk
None of these elements works in isolation. Discovery without granular guardrails leaves organizations knowing what exists but not controlling it. Harnesses without validation create a false sense of security. And governance focused only on access misses key variables that can introduce risk.
The organizations managing this well are the ones treating agent governance as an extension of existing IT and data governance, applying the same discipline used for foundational legacy systems rather than treating it as a bolt-on afterthought.
Given how quickly the technological landscape continues to shift, this comprehensive governed approach is what will keep agentic AI delivering value – and reduce the risk of agents going rogue.
We’ve featured the best AI tool.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
