The marketing pitch for enterprise AI’s autonomous agents has started to sound almost like a fairy tale: Hand one a task or objective, walk away, and it figures out the rest. It runs on its own, reasons through changing conditions, adapts as circumstances evolve, and delivers results before you think to ask. The promise of software that functions like a digital colleague has been seductive.
Amazon Web Services has an even more ambitious version of that vision in store. At AWS Summit on Wednesday, the company unveiled new agentic AI capabilities for its platform, aimed at everyday enterprise operations. The centerpiece is a set of updates to Amazon Quick, its workplace AI assistant for nondevelopers, that lets users create autonomous agents by describing them in plain language and deploying them in seconds with no code. Tell it to monitor overnight regulatory filings, compare them against company policies, and deliver an impact assessment by morning. AWS says the agent works continuously in the cloud and grows more effective over time, learning from interactions.
But the rest of the Summit announcements tell a stranger, more revealing story. The same company selling effortless autonomy is also shipping an arsenal of tools whose entire purpose is to watch those agents, second-guess them, and undo their work.
AWS unveiled a release-management capability for its DevOps Agent that vets AI-generated code for production readiness because, as the company frames it, coding agents now write at extraordinary speed while human review still crawls. It also introduced a tool named Zero Debt, built on the premise that the faster code is generated, the faster technical debt compounds—meaning cleanup must become continuous and autonomous, too. A new security capability begins every remediation in “learn mode” and graduates to autonomous enforcement only as confidence grows.
It is the design language of a company betting heavily on autonomy while quietly acknowledging how much could go wrong. So Fast Company put the obvious question to Swami Sivasubramanian, AWS’s vice president of agentic AI: If agents are ready for production, why does so much of this release exist to watch them, validate them, and roll them back? And if they require that much oversight, what does “autonomous” actually mean for AWS?
Sivasubramanian rejected the notion that the platform’s safeguards amount to an admission of weakness. In his view, they are the mechanism that allows organizations to trust agents at scale.
“Inserting deliberate friction into a process isn’t a sign of good governance or strong safeguards,” he tells Fast Company. “You could ask any business today, and they would trade that kind of friction in a heartbeat if they could do it safely and securely. The opportunity with AI is to replace manual friction with policy-driven controls that can operate at the speed and scale modern organizations require.”
He pointed to code review as an example of how AI is forcing companies to rethink one of software development’s oldest workflows. AI coding agents can now generate code far faster than organizations can review, test, and deploy it. That mismatch, Sivasubramanian argues, has become the next bottleneck.
Sivasubramanian says AWS DevOps Agent is meant to collapse several stages of the software development process into one continuous workflow. The agent does not just generate code; it can immediately review that code, create a testing environment, identify likely pipeline failures, and attempt fixes before a human has to step in. Rather than requiring an administrator to approve every action, organizations define policies, establish escalation thresholds, and monitor activity through audit and observability systems.
To Liz Miller, vice president and principal analyst at Constellation Research, that tension says less about a lack of confidence in autonomous agents than it does about where enterprise AI adoption actually stands today.
“In conversation after conversation with enterprise technology leaders, governance, risk, and accountability are always the leading concerns and constraints to advancing their AI and, more specifically, their agentic agendas,” Miller says. Viewed through that lens, AWS’s latest announcements are aimed as much at making agents usable within enterprises as at making them more capable.
“No matter how much someone wants to use AI, if the organization can’t de-risk the agents and the models, they won’t be allowed into production,” Miller says. Infrastructure updates such as AgentCore Harness, the managed runtime for agents, and AgentCore Policies may ultimately prove as important as the agents themselves.
Miller says the market is split between “two AI narratives”: one in which AI is “miraculous and capable of astounding things,” and another in which enterprises have to work through accountability, governance, cost, and business value.
“Yes, these models are incredible,” she adds. But business leaders still have to show how AI makes the business itself more incredible.
AWS’s intriguing definition of ‘autonomous’
The word has been worn smooth from overuse. Every booth at every conference now sells something “autonomous,” and much of it amounts to chatbots and robotic process automation in fresh packaging.
Sivasubramanian explained that reasoning and memory have improved enough over the past few months that execution is largely solved. The harder problem is making sure an agent can stay on the right path over time, continue making good decisions as conditions change, and remain aligned with the goals of the business. “For us, autonomy is defined by whether those actions remain trustworthy over the course of a long-running process,” he says.
That is why AWS is starting the security capability in learn mode rather than turning on full enforcement from the start. Sivasubramanian says the system can take on more autonomous action over time, as customers gain confidence and set the guardrails for which decisions it can make on its own.
In healthcare and other safety-critical domains, human control stays put. “What we’re building is not a binary model where humans disappear,” he says.” It’s a framework that allows organizations to decide where autonomy makes sense and where human oversight continues to add value.”
An incoming agentic AI Governance Nightmare?
The Quick assistant’s headline promise—describe an objective, get a working autonomous agent in under a minute—is also its sharpest risk. What stops a company from spinning up thousands of agents faster than it can govern them? Sivasubramanian’s answer is that governance should travel with the agent rather than depend on gates a human has to open.
“We want to make it easy for anyone to build agents and use agents in their existing systems in a way that is accurate, trustworthy, and secure,” he says. “A great agent is one thing that can help you deliver a great outcome for a customer, but it’s only one aspect.” The harder challenge, he argues, is everything surrounding the model: the infrastructure that allows it to run securely at scale, the data layer that provides the right context, and the tools that enable it to take action across systems.
The gap between what AI can do and what enterprises are willing to let it do has narrowed considerably, says Miller at Constellation Research. The headline-making failures, in her view, will belong to organizations that mistake automation for autonomy.
“We won’t see fully autonomous customer-facing roles being taken over by AI,” she tells Fast Company. “This isn’t to say the folly of firing your whole customer service or marketing team won’t happen—they will, and they will be headline-making disasters.”
The Problem Enterprise AI Still Hasn’t Solved
Every enterprise AI company’s unifying promise is getting agents into production faster. But the most-cited research suggests speed-to-deploy isn’t where enterprises are actually stuck. Gartner projects that more than 40% of agentic AI projects will be scrapped by the end of 2027, citing escalating costs, murky business value, and inadequate risk controls—none of which is fundamentally an infrastructure-speed problem.
So why is velocity the problem worth solving? Sivasubramanian argues that the skepticism resembles the doubts that accompanied earlier platform shifts, such as cloud computing. He points to customer examples, though not broad industry benchmarks. Mainframe modernization projects that historically stretched three or four years are now being completed in less than six months on some applications, he claims.
The deeper lesson, he argues, is that model intelligence is becoming less important than access to context. “One of the biggest lessons we’ve learned is that intelligence is no longer the primary bottleneck. Context is,” he says. “You can have a highly capable model, but if it doesn’t understand your systems, your policies, your data, your workflows, and the realities of your business, it’s limited in what it can accomplish.”
Still, across dozens of announcements, AWS provided no error rates, accuracy benchmarks, or time-in-production metrics for its autonomous claims.
Autonomy Doesn’t Eliminate Accountability
Every agentic AI announcement raises the same unresolved question: Who is accountable when something goes wrong? Software may take on more execution, but companies still own the outcomes. A security agent can trigger an outage. A business agent can make the wrong call. An AI-generated release can still break production.
Sivasubramanian acknowledges that the industry is still working through what that reality ultimately means. “With the pace technology is going today, it would be hard for anyone to look five years out and tell you with any certainty what should or should not be automated,” he says. “We don’t deploy patches instantaneously everywhere. If the agent encounters conditions that the developer set in advance, it will bring the developer in for review.”
He was also blunt about what automation can’t offload. “Humans approve fewer individual actions while remaining responsible for the system-level decisions that determine outcomes,” he says. “The approval surface shrinks to a few big priorities. The accountability doesn’t.”
If autonomy is genuinely ready, the guardrails are a courtesy. If it is not, they are the product. Either way, AWS appears to be making a claim that the company that wins the AI agent era may be the one everyone else has to route through.
