AWS ships Continuum and Context to fix what AI agents keep getting wrong

AWS shipped two new services at its Summit in New York on June 17, 2026 that target the two failure modes enterprise AI agents keep hitting in production: confident but wrong decisions, and code shipping faster than security can review it. Continuum and Context headlined agent-focused updates that brought Kiro to iOS, added Release Management to AWS DevOps Agent, expanded Amazon Bedrock AgentCore, and pulled Southwest Airlines in as a flagship customer.

Continuum is a security service built for the "agentic world" AWS now describes as the default threat model. It scans for code vulnerabilities, ranks them by business impact, tries to reproduce successful attacks in an isolated test environment, and only then suggests fixes. Context is a knowledge graph service that infers relationships between enterprise data, business rules, and domain knowledge, and exposes that layer to every agent in the organization. Both services start in modes that require human sign-off and graduate to enforcement as teams build trust. The full announcement is in AWS's own writeup of the Summit by Swami Sivasubramanian, AWS VP of agentic AI.

What Continuum does and why it is shipping now

Continuum is AWS's bet that the security stack has to become agentic before the rest of the stack does. The service runs the full vulnerability lifecycle, from discovery through prioritization, validation, and remediation, with continuous scanning layered on top of the existing backlog. The validation step is the interesting one. Instead of trusting static analysis or CVE feeds, Continuum tries to reproduce the attack in an isolated test environment. Findings that cannot be reproduced get downgraded. Findings that can get prioritized by business impact: is the affected component reachable, is it in production, what is the blast radius.

The new threat modeling companion tool does the same for design documents and source code, generating attack scenarios in an industry-standard format that existing security tools can ingest. AWS positioned this as the natural follow-up to specialized security models like Claude Mythos showing up in the wild. When attackers and defenders are both running the same model class, the side that automates the loop faster wins.

The timing is also clearly informed by AWS's own recent incidents. In February 2026, reports tied Amazon's AI coding tools to a 13-hour outage after the Kiro coding agent decided to delete and rebuild an environment. Amazon put an internal policy in place shortly after requiring experienced engineers to approve all AI-generated code before it ships. The Release Management feature announced alongside Continuum is the customer-facing version of that policy: a Release Readiness Review that checks every code change against production requirements, flags cross-repository dependencies, and runs the change in a production-like environment before it lands.

The context layer and the wrong-answer problem

Context addresses a different problem that is harder to point at: the agent gives an answer that is technically correct and still wrong, because the answer was based on a stale view of which table is authoritative or which source the company actually trusts. Knowledge graphs are not a new idea for this problem, but building one is usually a six-month professional services engagement. Context is AWS's answer to that gap. It automatically builds the graph from existing enterprise data, derives the relationships from databases, documents, emails, and chat messages, and layers in business rules and domain knowledge.

The interesting design choice is the feedback loop. Every query to Context records which sources delivered reliable results. Later agents benefit from earlier ones. In practice this means the graph gets more useful the more the organization uses it, which is the same compounding argument AWS made for agents in general: more interactions, more context, better outcomes, more trust, more work handed off.

Context is also positioned as the connective tissue between agents and the rest of the Amazon Quick, Kiro, and Bedrock AgentCore stack. Quick's new autonomous agents, Kiro's iOS app, and the DevOps Agent's Release Management feature all run better when they can ground their decisions in a Context graph that the rest of the organization has been training. The Bedrock Managed Knowledge Base, which is shipping as part of the AgentCore update, is the same idea wrapped in a different integration: ingestion, parsing, and retrieval for retrieval-augmented generation, with native connectors to S3, SharePoint, Confluence, and Google Drive.

The wider agent stack and what it signals

The rest of the Summit announcements read as a portfolio play rather than a single product launch. Kiro on iOS is the consumerization play: same coding agent, accessible from a phone, sessions running in a cloud environment, with identity and model settings syncing across the IDE, web, and mobile. The Bedrock AgentCore update is the platform play: more connectors, more security filters from partners like Check Point, Zscaler, Rubrik, Netskope, and SentinelOne, and a usage stat that AWS used to anchor the momentum argument. AWS's earlier AgentCore CLI push to make agent delivery a three-call operation is a useful reference point for how aggressive the platform-side consolidation has been over the last quarter. The number of tasks performed by agents on AgentCore has grown 15x in the last six months. The PGA Tour is writing tournament coverage 10x faster. Nasdaq, Visa, and Experian are scaling agents across the enterprise.

Southwest Airlines is the marquee customer story. The airline is moving from a largely on-premises footprint to a cloud-based, AI- and agent-enabled architecture on AWS by 2028, with more than 2,700 developers already using Kiro. Southwest is also adopting an AI-Driven Development Lifecycle, or AI-DLC, approach, where AI agents move development forward and engineering teams guide and validate outcomes. It is the most concrete large-enterprise commitment AWS has lined up to anchor the agent narrative for the rest of 2026.

The shape of the announcement is worth paying attention to on its own. Continuum answers the security review problem. Context answers the wrong-answer problem. Kiro on iOS answers the access problem. Release Management answers the production-shipping problem. AgentCore answers the integration problem. The portfolio is not a single product. It is an attempt to make AWS the place enterprises go when they want to put agents into production rather than into a pilot.

For teams already running on AWS, the practical next step is to look at the Bedrock AgentCore managed knowledge base and the Quick autonomous agents as the entry points, then layer Continuum and Context on top as the production hardening. For teams on other clouds, the comparison is now sharper: who is the AWS-equivalent of Continuum, and who is the AWS-equivalent of Context? The agent stack is moving from "can your vendor host a model" to "can your vendor host the whole loop." The Summit announcements are AWS's answer to that second question. For a broader look at how enterprise AI agent rollouts are actually structured, the enterprise AI resource page walks through the same set of decisions for production deployments.