Strands cut agent costs in half, and I published a book — Tech & AI Weekly, June 19 2026

Jun 19, 2026

Hi! Quieter week on my side, and I needed it. I spent it heads-down on research and shaping ideas for new content, including a sample I’ve been building on why agents fail in production and how to make them improve between runs instead. So this issue leans into exactly that: agents that hold up, cost less, and don’t fall over the moment a real user shows up. Here’s what mattered, curated so you don’t have to follow 20 sources.

AI & GenAI

Strands Agents shipped context management that cuts costs ~55%, plus a sandbox and Evals 1.0. One line, context_manager="auto", offloads big tool results and compresses old messages instead of dropping them, with compression firing at 85% context usage. On their code-investigation benchmark, cost dropped 55% while accuracy went from 68% to 98%. They also shipped Strands Shell (run agent commands in an isolated sandbox with allowlisted files and network) and Strands Evals 1.0 with chaos testing and red teaming. Read more
My take, in tutorial form: stop wasting tokens with the wrong agent memory. Right as the tooling ships automatic context handling, here’s the why underneath it. I break down conversation memory vs context, and how confusing the two quietly inflates your token bill. Read more
GLM-5.2 released full open weights and tops the open-model rankings. 753B parameters, MIT licensed (you can download it, run it on your own hardware, and fine-tune it, with no API gatekeeper that can pull it on you), a 1M token context window, and it leads the open-weights index at roughly a third of GPT-5.5’s price. One real downside noted in testing: it’s token-hungry. If you want a capable model that’s yours to keep, it’s worth a look. Read more
Hugging Face’s “is it agentic enough” benchmark found that agent skills help big models and hurt small ones. Their agent-eval harness measures the work an agent spends, not just the final answer. Giving models a CLI and packaged docs sped up large models, but small ones got worse: one (Qwen3-14B) dropped from 67% to 43% accuracy and tried to call the CLI as if it were a built-in tool. The lesson: test agent-facing tooling across model sizes, because the same feature can help one and break another. Read more
Microsoft announced Scout, an “always-on” agent with its own identity. Built on the open-source OpenClaw framework, it reads and writes local files, runs shell scripts, applies code patches, and spawns sub-agents. Microsoft’s answer to the obvious security questions: each Scout gets its own scoped Entra identity, credentials are redacted from logs, and sensitive actions need human sign-off. A clear sign of where the big platforms are pushing: from chatbots to autonomous, long-running agents. Read more

From AWS Summit New York

Busy week at AWS. The launches I’d look at if you build agents:

AWS Context: a knowledge graph your agents query at runtime. Announced in the keynote (it’s “coming soon”), it maps relationships across your scattered data into a governed graph, then gets smarter as agents use it: it watches which sources give correct results and which join paths work, and shares that across the org so one agent’s discovery helps the rest. Every query inherits the calling user’s IAM permissions, so it’s auditable by default. If you’ve hand-built a retrieval pipeline, this aims to curate the graph for you. Read more
Web Search for AgentCore went generally available. A fully managed tool that grounds agents in current web knowledge with zero data egress, the query stays inside your AWS environment and isn’t sent to an external search provider. My friend Axel Piérola wrote a sharp breakdown of the difference between this and the Browser tool: if your agent needs to know something current, use Web Search; if it needs to do something on a site, use Browser. It pairs naturally with the memory piece above, both are about feeding agents trustworthy context. Read more
Gemma 4 is now on Amazon Bedrock. Google’s open-weight family (Apache 2.0) in three sizes, all multimodal with a 256K context and native function calling. You call them through an OpenAI-compatible endpoint, so if you’re on the OpenAI SDK you switch by changing the base URL and model ID. More open options in your model layer. Read more
Amazon Quick added autonomous agents. They run continuously in the background, flagging stalled deals and drafting follow-ups, delivering overnight impact summaries on regulatory changes, handling purchase orders. You build one by describing what you want in plain language, set its autonomy level and guardrails, and it sharpens over time. Worth a look if you lose hours to repetitive triage. Read more

Cloud & DevOps

SpaceX bought Cursor for $60 billion. Another piece in Elon Musk’s AI empire, alongside X and Tesla, and the deal was big enough to push SpaceX past Amazon as the world’s fifth most valuable company after its Nasdaq listing. The stated reasoning ties to a future plan to put AI data centers in space. One to watch. Read more

From me this week

I wrote a book: Architect Your Agent’s Harness. It’s out on Kindle. I take one travel-booking agent, the kind that demos beautifully and breaks in production, and rebuild it in front of you so it stops inventing hotels that don’t exist, stops grabbing the wrong tool, and stops burning tokens in loops that never end. The whole argument: the prompt was never the problem. An agent is code, a harness around the model, and reliability is something you engineer into it layer by layer. By the end you don’t hope it works, you can prove it does. Get it here

Where to find me

SF Pitch Night at the AWS Builder Loft, Thursday June 25. I’m hosting, and if you’re in San Francisco you should come by. Six hand-picked AI teams pitch to founders, investors, and operators, with live audience voting to crown the winner. It’s at the AWS Builder Loft, 525 Market St, doors from 5:30 PM. Come say hi if you’re around. Details

I hope you enjoy this as much as I enjoy putting it together every week. It’s my little weekly de-stress moment. Let’s stay informed together and learn something new along the way.

See you next Friday,

Eli

Dev.to · LinkedIn · GitHub · Twitter/X · Instagram · YouTube