AWS spent its New York Summit keynote on the problem agent demos skip: getting work across the boundaries between your systems.
By Maribel Lopez, Lopez Research | June 2026
The agentic AI conversation has progressed through three stages over about two years, and the AWS New York Summit discussed aspects of the third wave.
The first stage was to define the concept of an AI agent and provide companies with tools to build one. The second, where most vendors have spent the past year, was building the scaffolding to run agents in production, such as testing, observability, identity, memory, and governance. AWS calls its version Bedrock AgentCore. Microsoft and Google have built their own platforms, and NVIDIA’s NemoClaw added a security layer for open agents. Phase two solves important parts of the agentic AI landscape, and the industry is still working through these problems. Dr. Swami Sivasubramanian, VP of Agentic AI at AWS, named directly: “too many agents are stuck between prototype and production.”
The third stage requires breaking down the data siloes and getting the various tools to work together. Sivasubramanian stated the problem plainly when he said, “The problem is not any single tool per se, it is the space between them.” It’s the fragments you need to finish a task that are scattered across Slack, email, a dashboard you haven’t opened in days, and a doc someone shared last week. Every login is, in his words, “another place where context goes to die.” And every time you’re the one connecting the dots, your momentum stalls.
Today, you can get AI agents to work within specific boundaries such as your CRM, ERP, and Talent applications. The challenge arises when you need to get agents to work across various applications and data sources. The data is in different formats, lives in many places, and the ways it can be shared or accessed vary by source. When an employee is trying to complete a task, it’s a human orchestration engine with permissions that allow them to gather, synthesize, and complete tasks across the varied landscape. Eliminating the friction is exactly where AI agents become genuinely useful and genuinely risky at the same time because it crosses your data boundaries on your behalf.
What made the problem statement credible at the AWS Summit wasn’t simply Amazon saying it. It was the buyer saying the same thing independently on stage. Lauren Woods, EVP and CIO of Southwest Airlines, described what the 2022 Winter Storm Elliott disruption taught her team: “It’s not just about systems working; it’s about how systems work together at speed and at scale.” Her systems did what they were designed to do. They just weren’t built to keep pace across the whole operation at once. That is the buyer-side version of the space-between-tools problem.
From Copilots to Cross-System Work
It helps to trace the path because the industry is solving these problems in sequence.
We started with AI chat assistants and copilots. We gave them chat windows, connected them to a few tools, and they answered one question and forgot it. They worked in isolation. The original copilots were, in effect, better search bars. Useful, but bounded. Swami’s version of the same point: the promise was intelligence, but “what we got was a slightly faster search bar, and faster search doesn’t compound, it flatlines.” A faster search bar solves one problem, but the real work requires moving between islands without a human acting as the orchestration layer. To get to the next level, we need AI agents that operate across all your systems with identity, security, observability, and governance in place.
Why Compounding Momentum Changes the ROI Math
AWS’s term for what that unlocks is “compounding momentum”. The value of an agent isn’t the first task it completes. It’s the slope of improvement after that. Swami’s framing: “context is what makes your agent’s 10th decision better than its first.” Describing the organizations pulling ahead, he added: “Every task that their agents complete makes the next one smarter.”
For a CIO, that reframes the business case. You aren’t buying a one-time productivity bump. You’re buying an asset whose return depends on accumulated context. It also means the cost of a stalled or siloed agent compounds in the other direction. Every month it isn’t crossing your systems is a month it isn’t getting smarter.
The plumbing for this is the knowledge graph. AWS describes Quick service as “an agentic search layer that works across your entire data estate,” and the company reports its internal semantic store already handles more than 1.8 million requests a day. AWS also announced a new managed service, AWS Context, that “automatically builds a knowledge graph from all your existing data.” Amazon’s goal is to create the connective tissue among raw data, knowledge bases, and business relationships and deliver a single governed context layer that agents query at runtime. The context layer should be able to operate with multiple models, allowing agents to provide a more accurate answer versus a confident but incorrect answer.
AWS Positioned Amazon Quick as an Overlay, Not a Walled Garden
AWS spent real stage time on Amazon Quick, and the strategic pitch was explicit. Swami framed the market as a false choice between agents trapped inside a “walled garden” that only see their own productivity suite, and open tools that reach across systems but bring no governance, identity boundaries, or control over where data goes. His claim: “This is a false choice. Quick doesn’t ask you to choose.” Quick is positioned as a governed overlay that works across systems you already run, including Slack, Google Drive, OneDrive, Snowflake, and Databricks, with each action carrying its own audit trail.
This is good positioning with real merit, and I’ve tested the functionality. It appears to work. Lauren Woods said she uses Quick “every single day,” and offered a concrete buyer outcome. The Southwest Airlines teams are “moving from looking at data after the fact to interacting with it in real time.” AWS also cited customer proof points, including GoDaddy saving a reported 15,000 hours of manual work annually.
But the questions that decide whether an overlay lands in the enterprise aren’t answered by a demo. They’re answered by economics and commitment:
- Licensing cost at full scale. An overlay that touches everything has to be priced so that “everything” is affordable across a large organization. This is the variable that quietly kills horizontal tools, and it’s the one buyers should model first.
- Durability of the investment. Enterprise buyers have watched products get announced but then not maintained. It’s a fair question to ask whether Amazon treats Quick as a long-lived product line or an evolving AI moment. If it gets enterprises genuinely engaged, it’s a strong business for AWS, which is a reason for optimism about its staying power.
- The adoption track record. Microsoft and Google have spent considerable effort getting copilots deployed broadly, and it’s been an uphill climb. To be fair, both have meaningfully advanced their offerings in recent months so that the pattern may be shifting. AWS enters a competitive landscape alongside other hyperscalers and solutions from model providers such as Anthropic’s Claude Cowork. Each solution from each vendor has its own merits and detriments, which also makes it hard to compare apples to apples.
None of that argues against Quick. It argues for evaluating it on what demos never show. For any solution, you need to consider the total cost at full deployment, whether it integrates with other tools you use, and whether you believe the vendor will still be investing in the product in three years. Building an agent is no longer the hard part. With the right tools and experimentation, you can build an AI agent that works. The connective tissue, such as memory, persistent context, the knowledge graph, and governed cross-system access, is what determines whether an agent finishes the job or stalls at the first boundary. The question isn’t whether agents can do the work. The question is whether your data, your identity, and your governance strategies are ready for software that crosses all of them on your behalf.