Kiro Crew, DynamoDB Vector Search, and Cheap Open-Weight Models

Top-Line Summary

AWS shipped Kiro Crew, a desktop agent orchestrator that imported Brett’s entire Claude workflow (skills, custom MCP servers, memory system) on first launch and burned half his monthly credits in a day. That import is the story, because a tool that absorbs months of accumulated workflow buys itself a switching cost nobody budgeted for. Elsewhere: DynamoDB got native vector search, GLM’s reasoning levels lie to you, and Chinese open-weight models are doing the same thing to model pricing that Chinese EVs did to car pricing.

Show Video

The “Pre-Show” Context

Short week in Canada. Brett had the Monday off and has been a day behind on everything since, Travers took the Tuesday and got an extra-long weekend out of it. Both of them are now waiting for the first CNE ad of the year, which is the sound of summer ending in Ontario. Brett’s plan for the changeover is to dig out the “winter is coming” t-shirt, which he prefers to wear in May specifically to annoy people.

The Engineering Rundown

• Kiro Crew (02:05)

  Kiro is a suite now: CLI, IDE, and Crew, which Brett describes as the technical person’s version of Amazon Quick Desktop. He installed it, it offered to import his existing setup, and it pulled in every Claude skill he has built over the last several months, his custom MCP servers, and his memory system through the Docker MCP Toolkit. That import is the part worth sitting with, because once a tool has absorbed your workflow you are not switching away from it for the same reason you do not switch IDEs. It spent the first hour insisting he was on the free tier, which turned out to be Crew inheriting credentials from the Kiro CLI, still logged in under his AWS Builder ID. He got real work done with it and half his monthly credits gone in a day on the $20 plan, but it is chatty rather than autonomous so far, and the LinkedIn pitch of waking up to a pile of finished work has not arrived yet.

• Security Hub Extended adds supply chain checks (13:04)

  Neither of us knew a Security Hub “extended plan” existed until this release added supply chain checks to it. From a first look it reads as the enterprise tier, with third-party tool integrations and coverage that reaches outside AWS, which probably makes sense for a multinational and very little sense for a small shop. Brett has it on his list to work out properly, because a couple of his larger customers might genuinely want it. The timing is funny given Kiro Crew ships its own extensions marketplace, where Brett already spotted one extension from an author who is not Kiro. Nothing bad implied about that author, but new does not mean you switch on every bell and whistle before you know where the code came from.

• GPT-Sol, Terra, and Luna get long context on Bedrock (15:20)

  A million tokens of context is now available on Bedrock for the Sol, Terra, and Luna models. Having a million tokens does not mean you should fill a million tokens, and both of us have watched output quality sag as context climbs. It is there if you need to load something genuinely large, and it pairs with the web search release below as a way to feed grounded material into an agent instead of stuffing the window with everything you own.

• Token caching (16:01)

  Brett has been circling token caching lately as a cost lever and wanted a sanity check on what it actually is. It is what it says: the tokens you send up and get back get cached and reused across prompts, so shipping an entire repo to a model does not cost full price on every turn of a session. Codex and Claude do it automatically through their APIs, and the assumption is Bedrock does the same rather than making you switch it on, though neither of us has confirmed that. Worth verifying before you build a cost model around it.

• DynamoDB now supports real-time vector search (17:52)

  Embeddings can now live in DynamoDB natively, up to 4,000 dimensions, with the usual distance options like Euclidean and dot product. Travers is planning to move his second brain off its current store, and Brett is looking at his Aurora Serverless setup running pgvector and doing the same arithmetic. The specific pain that pushes him is scale-to-zero: he scales Aurora down to control cost and then eats occasional timeouts on the way back up, which is not a problem DynamoDB has. There is no plausible world where Aurora comes out cheaper than DynamoDB for what either of us is doing here. The broader point is that the list of AWS data stores where you cannot use vectors is now shorter than the list where you can.

• Web search on Amazon Bedrock for model grounding (20:53)

  Bedrock agents can now search the web and pull back semantically relevant snippets rather than whole pages. The useful version of this is not filling your million-token window faster, it is pointing an agent at a known trusted source and having it validate a result before it hands you an answer. Steering beats stuffing.

• CloudFormation IaC MCP server (21:56)

  Roughly an LSP for CloudFormation: a type checker for your templates, with two validation tools wrapping cfn-lint and cfn-guard. Both of us had the same reaction, which is that any competent agent already writes a pile of CloudFormation without help, and in our world the linting and guard checks are already wired into CI pre-deploy. It feels a bit like building an MCP server because you are supposed to build MCP servers. That said, if you have thought about adding template validation and never found the time, this is a cheaper path to it than building the pipeline yourself, and there is nothing wrong with catching a bad template before it deploys.

• AWS MCP server sprawl (24:59)

  There are more than fifty of these now, including per-service ones like the CloudWatch server with its nineteen tools, which Travers spent a day fighting while trying to get Bedrock AgentCore to call it from a chatbot. Brett’s framing is that MCPs are APIs for AIs, and when your platform is 200+ services that are all APIs with a GUI bolted on, fifty wrappers is not a shocking number. The per-service split is defensible too, because there is no reason to load ElastiCache tooling into an agent that will never touch ElastiCache. We are putting an AWS MCP server walkthrough on the list for a future episode.

• A $10-a-month cloud dev box (28:09)

  Travers kept OOM-ing his laptop by spawning too many simulation jobs at once, so he moved his agent sessions onto an Ubuntu EC2 instance on a spot ASG. It runs tmux and mosh so sessions survive a disconnect, a user data script refreshes the Cloudflare DNS record on boot, Caddy fronts it in a private subnet, and a short bash script pulls credentials and connects from whatever device he is on, including Termux on his phone. A week in, it is tracking to about ten dollars a month. Brett has landed in the same place from the other direction, running Ghostty on his laptop as the host and connecting from an iPad over Tailscale with mosh, which meant an afternoon on the porch tasking Claude from a mini keyboard. He also ran his own laptop out of memory the same day and had to go argue with Activity Monitor about it.

• Apple pricing and the return of the mainframe (33:24)

  Brett wants a Mac mini with more headroom and cannot stomach the current prices or lead times, which are up roughly 30% across the Apple line, with a spec’d-out MacBook Pro around ten grand. His theory on why you cannot get a Mac mini is every YouTuber building a ten-mini cluster to run a model that would be fine on a phone. There is speculation of hardware-as-a-service, where you subscribe rather than own, which businesses have done with leases forever and which feels strange as an individual. Meanwhile the dev box conversation has both of us running remote terminals into a central machine, so we have reinvented the mainframe and then complained about leasing.

• Kanban is back (39:43)

  Brett’s kanban project came back to life as the first real workload for Kiro Crew. He handed it a backlog of small polish items he had been ignoring for months, told it to work on a branch so anything broken could be reverted, and it closed a stack of them while he did other work. None of them were things he needed, which is exactly why they had sat there.

• Polis, the city simulator that turned into a model benchmark (40:45)

  Four cities, each run by a different model, each able to produce only two of the three resources it needs to survive, which forces them to trade. Trades are agreements rather than contracts, so a model can promise ten materials, take your eight food, and simply not deliver. Opus does this constantly, and Brett has watched the tit-for-tat play out in the generated highlight reels: one city shorts another, the other retaliates, it escalates until somebody is close to dying, and then they both back off. The more useful finding came out of a tangent. He started counting tokens per turn out of curiosity and found GLM burning several times more than everyone else on the same “low” reasoning setting, which is how he learned that low is not low for GLM, and you have to set reasoning to none to actually turn it off. Before that correction GLM looked like it was winning everything by a mile. After it, Opus wins every time, by cheating.

• Open-weight models go mainstream (49:23)

  When the CBC runs two AI explainers in a week, the terminology has escaped the industry. This one walks a general audience through what an open-weight model is, using a recipe analogy that is genuinely good: a recipe gives you the ingredients and the measurements so you can adjust them, a closed model hands you the cake. The other thing the article catches is North American AI CEOs discovering a sudden enthusiasm for restricting open models, having already scraped the entire internet, spent a fortune, and struggled to show a return. Regulation looks a lot more appealing from the top of the pile.

• Four of the top five models on OpenRouter are Chinese (51:28)

  Two different DeepSeek V4 Flash builds, MiMo v2.5, and Hy3, with GPT-5.6 Luna as the only non-Chinese entry and climbing fast after its price cut. On artificial analysis numbers, Fable 5 runs about $3.15 per intelligence index task, Kimi K3 delivers broadly comparable performance for around 86 cents, and DeepSeek V4 Flash does it for roughly three cents. The comparison Brett keeps coming back to is Chinese EVs. Geopolitics is a real thing, but a person who needs a car looks at $40,000 next to $20,000 and asks what the extra twenty grand buys, and if they cannot answer that, they buy the cheaper car.

• OpenAI cuts prices on Terra and Luna (52:37)

  Up to 80% off, and notably not on Sol, the flagship. The stated reason is training efficiency. The likelier reason is a model that does the same work for three cents. What Brett would like to see next is the same treatment from the other big labs on last-generation models: if Fable is expensive because it is the frontier, then Opus 4.7 should get cheaper as it stops being the frontier. Frontier keeps its cachet and its price, everything below it becomes a utility.

• AI summaries are draining traffic from source sites (58:16)

  The article talks to a recipe site owner whose traffic fell off a cliff, for the obvious reason that the recipe is now sitting at the top of the search results. Google has an awkward problem underneath this, because their $224.5 billion search advertising business depends on sending people somewhere, and it is not clear what an ad inside an AI overview looks like beyond charging even more for the top slot. On the publishing side, SEO turns into building for agents: attribution links at the bottom of a chatbot answer may be where traffic comes from now, which makes llms.txt look less optional than it did six months ago.

• Reddit is unhappy about AI overviews (01:03:12)

  Reddit is positioning itself as the human internet, the place for deep discussions and passionate debates, which is not the Reddit either of us reads. They are reportedly considering cancelling a Google licensing deal worth around $60 million a year, on the grounds that AI overviews are cutting traffic to source sites by nearly half. Cancelling now does not help. The training data is already gone.

• Sam Altman’s parenting advice (01:04:47)

  The suggestion is that you connect your calendar to your kids’ calendars and generate a morning podcast that briefs you on what your children are doing. For a company of that size and importance you would expect enough media training to stop this one before it left the building. The best comment Brett saw in response was the simplest: talk to your kids.

• Custom podcasts that are actually useful (01:06:07)

  Brett flipped the idea and it turned out to be the most practical thing he tried all week. He is studying for a cert, built a skill called rap session that quizzes him and generates flash cards as he works through material, and at the end of a session it summarises the session into a podcast he can listen to later. The first attempt came out in Apple’s built-in voices, which sound like 2009 and made him laugh out loud, so he wired in an OpenAI key instead. Nineteen minutes of two-host audio cost nineteen cents. The turn-taking is still not human (no interruptions, no overlap, a beat of latency between speakers), but as a way to re-consume a study session it works, and pairing it with flash cards gets you close to a spaced repetition setup on demand. Both of us think the training and exam-prep sites should be nervous.

• Meta ships Muse Code and Muse Spark 1.2 (01:12:31)

  Muse Code is a terminal coding agent with simple agent loops and async background agents that feed the main agent, plus a local event log of everything it did so it can resume after a crash or interruption. That log is also an audit trail, and Kiro Crew appears to do the same thing. Muse Spark started as a model purpose-built for Meta’s own products and 1.2 adds long-horizon coding, which suggests it is aimed outside Meta now. The model and the harness look co-trained, so whatever numbers Muse Spark 1.2 posts probably only hold when you run it inside Muse Code, which makes “should I use one over the other” the wrong question.

• AWS is reportedly retiring the networking specialty exam (01:18:11)

  Unconfirmed, but Brett heard it today and neither of us will miss it. Along with the machine learning specialty, it was one of the two exams that could genuinely make you consider a career change, and Travers closed the laptop after his first attempt wondering what had just happened to him. The complaint is specificity: pages of BGP internals when the actual AWS task is handing a Transit Gateway an ASN. Nine and a half times out of ten on AWS you do not need to know the inner workings of BGP. Expect it back eventually as a foundational exam that is not foundational, same as the ML one.

Off-the-Clock Recommendations

• King of the Hill, new season. Travers’ standout of the week, and yes, it is the one where the guys stand around at the fence.
• Spider-Man, the new one. Decent, and it feels more like a Spider-Man movie than the other Tom Holland entries.
• Furious (Hulu/Disney). Murder mystery, a few episodes in, with subject matter that will make some people uncomfortable. Scoot McNairy plays a local cop and Emmy Rossum is in it, whose name took the two of us three attempts to land on air.
• Halt and Catch Fire (2014). Brett called it Control-Alt-Delete on the show, but seeing Scoot McNairy in Furious made him want to rewatch the one about his era of computing.
• Battlefield. Brett’s plan for the evening was grinding weeklies to unlock gear, instead of listening to his own study podcast.