NEW YORK, Feb 12, 2026, 16:56 (EST)
- Kyndryl is rolling out new “policy-as-code” controls, aimed specifically at agentic AI workflows operating in regulated settings.
- PYMNTS is flagging what it calls an “accountability gap” now that AI agents are moving past the copilot stage and starting to make actual decisions.
- CIO.com is pushing for an “agent control plane” to enforce strict guardrails on probabilistic AI.
Kyndryl rolled out a new “policy-as-code” feature on Wednesday, aiming to enforce organizational and regulatory rules directly in software for agentic AI—autonomous systems that handle tasks in fields like finance, supply chains, and government. The company says it translates complex business and compliance rules into machine-readable instructions, tracking every action taken by these AI agents and stopping any unauthorized moves. “The approach provides the structure customers need” to work with agentic AI, according to Ismail Amla, senior vice president at Kyndryl Consult. Regulatory and compliance worries remain a top obstacle for 31% of Kyndryl’s customers looking to ramp up tech spending, the firm said, noting it orchestrates close to 190 million automations monthly. PR Newswire
Companies are pushing AI agents past routine inbox chores and into handling real tasks, setting off new controls, PYMNTS reported Thursday. These systems are taking on more responsibility — even making decisions — while oversight struggles to keep up, the outlet noted. That “accountability gap” is starting to show as tools like Microsoft’s Agent 365 and Salesforce dashboards attempt to rein in what PYMNTS calls “agent sprawl.” Harvard Business Review, according to PYMNTS, is pressing firms to hire “agent manager” roles, treating AI agents like staff. PYMNTS.com
In a Thursday column for CIO.com, Wipro consultant Hari Om Garg took up the architecture angle, pushing for an “agent control plane”—essentially a deterministic software layer that grabs agent output before it hits production. Garg pointed to spend caps and vendor checks as examples of guardrails, with exceptions routed to people. “We cannot fire an algorithm,” he wrote. “Trust is not a feeling; it is a code module.” CIO
Agentic AI generally leans on large language models (LLMs)—AI built to guess the next word—woven into workflows that can run tools, spark actions, and pursue objectives. All of that adds utility, but it complicates oversight compared to standard automation, which is predictable and consistent.
Kyndryl wants agent actions corralled—kept inside set boundaries. According to the company, its policies enforce “deterministic execution.” That means an agent sticks to actions spelled out by established rules. Compliance and oversight? Dashboards and audit logs track it all, making every move visible for review.
Kyndryl claims its guardrails help rein in the operational fallout from hallucinations—when AI spits out convincing but incorrect responses—by preventing unpredictable or out-of-bounds actions as agents work through a process. The company pitches this as a way to shift agent deployments beyond pilots and into production, even in mission-critical settings, while keeping a firm handle on control.
Still, code-based policy isn’t bulletproof. When rules get revised or exceptions accumulate, agents may veer off course as data or models evolve. Controls set too broadly risk blocking legitimate work or letting mistakes slip by, which can bury teams in a mess of alerts and approval requests.
Kyndryl plans to lean on its Kyndryl Consult business to support clients building and managing agentic AI setups that meet company oversight and compliance requirements. The infrastructure giant serves thousands across 60-plus countries, as it works to stay relevant while AI pushes deeper into day-to-day business operations.
