AI Agents Need Controls Before They Need More Infrastructure

AI agents are already in production. Enterprises are giving them access to cloud environments, internal data, software tools, and workflows while governance systems struggle to keep pace. The difficult question is no longer whether agents can perform tasks, but whether organizations can control what happens when those tasks connect.

That pressure grows with every new agent. Felix Van de Maele, founder and CEO of Collibra, says enterprises running more than a handful of agents face risk, cost, and oversight burdens that grow exponentially with each addition. In a conversation with VentureBeat founder and CEO Matt Marshall at VentureBeat’s AI Week event in New York, Van de Maele traced most agent failures to the business context those systems receive and called for deterministic controls before agent counts rise further.

The cloud is where those concerns become concrete. The 2026 State of AI in Cloud Development report, conducted by Dimensional Research for LocalStack, surveyed more than 300 developers and team managers at organizations with 1,000 to more than 10,000 employees. The survey found that 79% give AI agents access to cloud environments for developing, testing, or staging code at least occasionally, while 18% said agents always have access.

That access comes with a remarkable lack of comfort: 97% expressed concerns about leaked credentials, misconfigurations, rising cloud costs, and failures in production. Yet 26% said agents can access production environments, and 55% said agents can access staging. Apparently, anxiety and access permissions can coexist quite peacefully.

The same survey shows why validation is becoming the bottleneck. Sixty-five percent of respondents generate at least twice as much code since adopting AI, but developers estimated that only 61% of AI-generated code reaches production. Twenty-six percent of developers said AI-generated code has more bugs than code created by humans, compared with 6% of development team managers.

“AI is generating code faster than teams can confidently ship it, making testing and validation the new bottleneck,” said Waldemar Hummer, LocalStack co-founder and CTO. He said LocalStack provides fast, safe, isolated cloud environments for developers and agents to iterate without the risks and costs of traditional cloud infrastructure.

More agents do not guarantee better results

Agent performance also depends on the machinery surrounding the model. A study that held five models fixed across three production harnesses found that a coding agent’s harness could triple its cost without making it more accurate. The vendor’s own harness bought no edge, and the only effect that beat the noise was a harness losing tasks.

That finding undercuts the easy assumption that better orchestration automatically produces better work. Standard multi-agent pipelines break down when complex problems cannot be divided into subtasks in advance, while Self-Organizing Agent Teams learn reusable collaboration strategies from past transcripts.

Even accurate agents can create incorrect outcomes when they work together. One agent’s output becomes another agent’s context; information can be lost or misinterpreted during handoffs, shared state can drift, and agents can enter loops or deadlocks. When several agents rely on the same underlying model, they may reinforce the same mistakes instead of catching them.

Tool selection adds another cost layer. Arcade.dev says its new Tool Recommendation system cuts input token costs by 37%, while emphasizing a layer that combines discovery with execution controls across customers’ chosen AI systems. Falling token prices have encouraged enterprises to expect lower AI bills, but agents are breaking that assumption as tool calls, verification, integration work, and oversight accumulate.

Platform adoption is not uniform either. VentureBeat Intelligence found in its September survey on agent orchestration that twice as many companies use OpenAI’s agent orchestration platform as Anthropic’s, even though Anthropic may lead OpenAI on other indicators. Ramp reports that more businesses now pay for Anthropic than OpenAI, a reminder that platform use, spending, and production value measure different things.

Governance must follow the workflow

AI governance cannot stop at a list of approved models. An employee asking an assistant to summarize a meeting transcript creates a different risk from an agent that can modify customer records or initiate a transaction. Governance teams need visibility into how employees use software and where AI-enabled activity creates process or compliance risks.

That means mapping ownership, permissions, approval points, and human involvement before adding autonomous capabilities. An agent that pulls from several internal applications must respect each source’s access restrictions; granting broad data access simply because an agent needs context creates unnecessary exposure. Higher-consequence actions need tighter controls.

Organizations also need better measures than licenses issued, active users, prompts submitted, or tasks completed. Human overrides, failed transactions, unauthorized access attempts, model errors, workflow exceptions, verification costs, integration expenses, and error-correction work reveal whether deployment is delivering value or just producing activity.

Several tools point toward that accountability layer. CTGT released Concordance, an open source tool for tracing why agents make inconsistent decisions. In one process, CTGT replaced identifiers with stable labels, sorted the list, isolated each test in a fresh organizational environment, and produced the same 34 obligations across five subsequent runs.

Cohere’s Model Vault encrypts AI inference so even Cohere cannot see enterprise customers’ data, with plans to open source the full serving stack so independent auditors can validate that it does not log, export, or leak data. These approaches address different problems, but both recognize the same constraint: enterprise AI needs evidence, not demos.

The next phase may depend less on how many tools a company deploys than on how well it controls the ones already running. Models change, integrations change, and access requirements change with them. AI agents can move through cloud environments and business workflows at impressive speed; without clear ownership, reliable infrastructure, and measurable outcomes, they can also move confusion at scale.

Stay connected via Google News

Follow us for the latest AI News