Most roundups just list what got announced. That tells you what happened but not what to do about it. June had a lot of model news, and one infrastructure announcement at AWS Summit New York that changes what it actually takes to get an agent into production. Here’s what happened and what it means if you’re building this stuff.

May was about the model catching up – Opus 4.8 on Bedrock, Managed Agents adding self-hosted sandboxes, Anthropic’s $65B raise. June was about the infrastructure around the model catching up too. Once both pieces are handled, the question that’s left is the one most teams skip past: which architecture actually fits the problem you’re solving. That’s where this piece ends up.

AWS Summit New York: the AgentCore harness hits GA

AWS Summit New York happened June 17 at the Javits Center. The announcement worth paying attention to wasn’t the headline one – it was the general availability of the Amazon Bedrock AgentCore harness. If your team has been building the orchestration loop, tool execution, memory, and error recovery by hand for every agent, read this part twice.

The harness is exactly what it sounds like: the scaffolding around the model. You define the agent in config – which model, which tools, which skills, what instructions – and AgentCore runs the loop for you. Two API calls, CreateHarness and InvokeHarness, get you a running agent with an isolated session, memory across turns, a filesystem, and web browsing. What used to take weeks to build now takes minutes. You can also swap model providers mid-session without touching the agent logic, so you could plan with one model and write code with another.

AWS also announced AWS Context, which builds a knowledge graph from your existing data and governs what each agent can see, and AWS Continuum (now in gated preview), which handles vulnerability discovery and remediation at machine speed. Both matter because agent tasks performed on Bedrock AgentCore grew 15x over six months, according to AWS, and that kind of growth is where thin data governance and weak security posture start to show. We wrote more on this in our Summit recap.

The short version: the harness used to be the hard part of building an agent. It isn’t anymore.

The rest of June’s frontier model news

Every major lab shipped something in June, and most of it pointed toward agents that can act on their own rather than just respond.

Anthropic had a busy month. Early on, they released Claude Fable 5 and a new tier above Opus called Claude Mythos 5, with a 1-million-token context window and always-on adaptive thinking. Within a week, a US government export-control directive forced Anthropic to pull access to both models. Fable 5 came back globally on July 1 with updated safeguards; Mythos 5 access returned for a set of US organizations. Then on June 30, Anthropic shipped Claude Sonnet 5 – close to Opus-level performance for coding and agentic work at Sonnet pricing – along with Claude Science, aimed at research and drug discovery. They also opened an office in Seoul on June 17 and signed a deal with California on June 29 to give state agencies discounted Claude access and training.

Microsoft used its Build conference to push its own models – seven new MAI models, including MAI-Code-1-Flash, which was trained directly against the GitHub Copilot harness it runs in production rather than distilled from another model. It’s rolling out across Copilot plans now, and Microsoft says it beats Claude Haiku 4.5 on price-to-performance for coding.

Google shipped Gemini 3.5 Flash with computer use built in, so agents can see and act across desktop, mobile, and browser, and Gemma 4 12B, a small model meant to run agents locally on a laptop. OpenAI expanded Codex with plugins connecting it to 62 business apps and 110 pre-built skills, plus a preview called Codex Sites for building internal tools from a prompt – clearly aimed at the non-developer share of its user base, which OpenAI says is growing three times faster than developer usage.

Picking the pattern is still the hard part

A harness that assembles the orchestration loop for you doesn’t tell you which loop to build. It will run a single agent with a couple of tools just as easily as a five-level hierarchy of subagents, and picking the wrong one is how you end up with something that’s technically live and still doesn’t work well.

Bala Balaiah, our Sr. Data Architect, wrote about exactly this in Choosing the Right Agentic AI Pattern for GenAI Implementation. He lays out 13 patterns across three categories – deterministic, dynamic orchestration, and iterative – and the line that sticks with me is: “most AI implementations don’t fail because the model was wrong. They fail because the architecture was.”

That was true when Bala wrote it in April, but it matters more now. True to the pace of innovation since April, picking a pattern and building the orchestration to support it were the same project, you couldn’t really separate the two. Now that the harness handles the orchestration, those are two different decisions. You can stand up any of Bala’s 13 patterns on a production harness in an afternoon, which means the pattern you pick is doing more of the work than it used to.

A single-agent setup on a fully governed harness is still wrong for a task that needs dynamic routing across specialists. A swarm pattern is still slow and expensive for something a sequential pipeline handles fine.

Driving Towards Results

If you’re building agents right now: the models are good enough to trust with longer, multi-step work. The harness will handle orchestration, memory, and error recovery so you don’t have to build it yourself. What’s left is picking the right pattern for the problem, and that’s not something any of these announcements can do for you.

That third piece is where most of our client work sits right now. We’re looking at how AWS Context and the AgentCore harness change the build sequence for projects that used to get stuck on custom orchestration, and pairing that with Bala’s framework to figure out the right pattern up front instead of after something breaks in production.

Bottom line

June gave us a harness worth building on and a set of models worth trusting. Neither one tells you what to build. That part’s still on us to assist our clients with.

Alec MacEachern is Vice President of AI at UTurn Data Solutions, an AWS Premier Tier Services Consulting Partner based in Chicago. Over the past decade, he has held roles at NVIDIA, AWS, and Microsoft, helping organizations design, build, and scale AI solutions across a wide range of industries and platforms. Today, Alec brings that cross-platform experience to helping enterprises navigate cloud migration, modernize data foundations, and adopt production-ready generative and agentic AI solutions.