The biggest announcement at AWS Summit New York yesterday was not a new model. It was a knowledge graph. I laughed when I read the headline, because it is the most boring-sounding thing that might actually matter.
AWS Context, launched June 17, maps an organization’s databases, documents, Slack messages, emails, and CRM records into a single graph that any agent can query. It learns which data sources produce correct results, which paths get used, and which business rules apply. Every agent benefits from what a single query discovers.
That is an infrastructure bet. Not a model bet. And it tells you exactly where the bottleneck has been all along.
The compounding loop AWS described out loud
Swami Sivasubramanian, AWS’s VP of Agentic AI, said something on stage that business leaders should pay close attention to: “The more you use agents, the more you get done. More interactions give agents more context. More context leads to better outcomes. Those better outcomes increase how much you trust them. The more trust you have, the more work you hand off.”
That loop only works when there is a context layer underneath it. Without one, every agent starts from zero. Every time. It checks the same tables, asks the same clarifying questions, guesses wrong about which data source is authoritative. You get an expensive chatbot, not a system that compounds.
AWS Context is the infrastructure that makes the loop real. It stores the metadata in Iceberg format on S3 Tables, plugs into existing tools, requires no retrieval pipeline to build. Governance is built in so agents only access what they should. And as agents interact with it, the graph gets better.
The strategic move here is not the technology. It is the admission: the model was never the problem. The plumbing was.
The numbers behind the bet
AWS shared several data points that tell the same story:
AgentCore, their platform for moving agents from proof-of-concept to production, saw tasks performed by agents grow 15x in the past six months. Nasdaq, Visa, and Experian are scaling agents through it. The PGA Tour writes tournament coverage 10x faster.
Dhan, a fintech unicorn in India, used Kiro (AWS’s coding agent) to build a new charting platform with a single engineer in eight weeks. The original estimate was 12 to 24 months with a dozen people. That is not a productivity improvement. That is a structural change in what one person can ship.
Southwest Airlines announced a full partnership with AWS, putting 2,700 developers on Kiro and transitioning to a cloud-based, AI-enabled architecture by 2028. They are adopting what they call an “AI-Driven Development Lifecycle” where agents move development forward while engineering teams guide and validate outcomes.
None of this is a pilot. Southwest is restructuring its entire dev org around this. Dhan replaced a year of roadmap with eight weeks and one engineer. These are bets companies do not walk back.
What this means for teams still evaluating
If you are still deciding which AI vendor to pick, you are solving the wrong problem. AWS just productized the layer between your data and your agents. Google did something similar with Vertex AI Agent Builder. Anthropic has been building toward it with tool use and MCP. Every major infrastructure provider landed on the same conclusion independently: the model alone does not produce business results. The wiring between the model and your data does.
This is the gap between a demo and a deployment. The demo works because someone hand-fed it the right context. Production fails because nobody built the system that feeds context automatically, at scale, with governance.
AWS Context is one version of that system. Yours might look different. But if you do not have one, your agents will keep hitting the same ceiling — smart enough to answer any question, and completely ignorant of your business.
The question nobody on your team is asking
Most AI conversations inside companies focus on capability. Which model is best? What can it do? How does it score on benchmarks? I sit in on these calls. Nobody asks the question that actually matters.
Does this agent know where our data lives, which source is authoritative, and what business rules apply to this decision?
That is not a technology question. It is an operations question — mapping data sources, defining access policies, deciding which system is the source of truth for which decisions. Boring work. The work that makes everything else function.
AWS just built a managed service for the job most companies have been avoiding. The model was ready months ago. The question was always whether you would do the plumbing. Now there is less excuse not to.