Pi Agent, Claude Design, and a $30K Bedrock Bill

Top-Line Summary

The headline this week is a $30,000 surprise: AWS Cost Anomaly Detection silently ignores Marketplace charges, and since Claude on Bedrock bills through Marketplace, one team watched their invoice climb with no alert firing. Around that, Brett and Travers spend most of the episode on agent tooling getting real — the Pi coding agent, building UIs from a screenshot with Claude Design, and what 1Password learned refactoring a multi-million-line Go monolith with agents.

Show Video

The “Pre-Show” Context

After a 48-hour blast of summer over the long weekend, the weather swung back to cold, so the timing for sitting inside and talking cloud news worked out. The two also compared notes on the LinkedIn livestream: nobody ever comments live, but LinkedIn pulls the most view time after the fact, probably just because it is the biggest audience either host has.

The Engineering Rundown

• EventBridge Scheduler Adds 619 API Actions (01:49)

  AWS added 619 new SDK API actions and 13 service integrations to EventBridge Scheduler, including Lambda Managed Instances. Brett’s first reaction was deadpan: only 619? Both hosts treated it as a quality-of-life update rather than a headline, but the real point is the pattern. EventBridge Scheduler now lets you pipe events directly to services instead of gluing things together with Lambda, the same way Step Functions and AWS Config managed rules evolved from a handful of integrations to covering nearly everything. If you are building event-driven solutions on AWS, this is where you should start, and writing Lambda glue should be the exception you reach for only when something is genuinely unique.

• Secrets Manager Agent Gets Pre-Fetching and IAM Role Assumption (04:40)

  Secrets Manager Agent picked up pre-fetching and IAM role assumption. The sweet spot is EKS workloads hitting Secrets Manager: batching requests cuts API costs, and you can pull secrets at task launch instead of on every call. Travers sees it as a straightforward efficiency win for anyone on EKS, and possibly managed ECS. They kicked around whether a Lambda function could use it to cache credentials, but landed on no — Lambdas already pull secrets from environment variables at cold start and live less than 15 minutes, so the caching layer does not buy you much there. Brett’s framing: why build caching from scratch when the service does it for you?

• Personal Projects: Travers’s LLM Harness, a Board Game, and a New Language (07:33)

  Travers is still chipping away at his own LLM agent harness, a project he has had going since January, and his takeaway is that writing good software solo is genuinely hard compared to having a team of experienced developers behind you. His board game is into rules tweaks and playtesting, with something to show next week. And in classic shiny-object fashion, he started designing his own programming language, which he fully expects to abandon by next week. Brett’s response was a story about a crew of Perl die-hards who once half-seriously planned to write an entire operating system in Perl, complete with a CN Tower launch banner.

• Personal Projects: An nginx CVE Ate the Long Weekend (09:28)

  Brett’s long weekend got eaten by an nginx CVE — the one allowing crafted requests that could cause denial of service or potentially remote code execution. It was serious enough that one of his customers issued a blanket change record for all of production: if your application used the affected component, you patched immediately. The patching itself went fine, but going back into older projects surfaced a pile of past decisions that no longer made sense, and Brett ended up creating roughly a dozen cleanup cards for himself. His honest read: the CVE generated a wave of follow-on work as he dealt with the decisions he made the first time around.

• The Pi Agent, OpenCode, and GLM 5.1 (12:26)

  Brett spent real time with the Pi agent and came away impressed, with caveats. It does not support MCP out of the box, and his attempt to have Pi port his Kanban MCP into itself went badly, so he switched to Pi’s extension model, which clicked. He built extensions for OpenBrain and Kanban, wired Pi into OpenCode to try different models, and found Qwen frustrating before GLM 5.1 turned it into a night-and-day experience: more expensive, but it just did the work. The standout is raw speed — Pi is fast enough that you can’t read the output as it streams, which no model he tried elsewhere could match. The catch is that Pi runs in YOLO mode all the time with no confirmation prompts, and Brett was candid that this is how people get themselves into deep trouble; he is now thinking about per-project Docker containers, scoped credentials, or even a remote EC2 dev box rather than running all-gas-no-brakes agents on a laptop that touches dozens of AWS accounts.

• Gemini 3.5 Flash and the Gemini CLI Shutdown (22:35)

  Google announced Gemini 3.5 Flash, roughly three times faster than 3.1 at around 300 tokens per second, and it jumped from the bottom third to the top third of the Artificial Analysis Intelligence Index, landing between GPT-5.4 and Kimi. The benchmarks point squarely at agentic tool use and coding. The catch is price: it is more than five times the cost of Gemini 3 Flash, and reportedly at or above Gemini 3.1 Pro, since it is a reasoning model. The hosts also flagged that Google is shuttering the Gemini CLI and folding it into Antigravity, a move they are not thrilled about, since the open-source Gemini CLI was a useful reference for how these tools are built, and walking away from a CLI runs against where everyone else is heading.

• Claude Design and the Impeccable Skill (25:50)

  Travers brought up the Impeccable skill — it breaks a project into design-language elements and hooks into Chrome so you can highlight elements on a live page and submit targeted feedback. That set up Brett’s story of the week: asked during a demo whether the plain-HTML chat UI on a Bedrock agent demo could look better, he tried Claude Design for the first time. He dropped in a screenshot taken during the call, answered a few multiple-choice prompts, and got back static HTML pages ready for S3 — and the share-to-Claude-Code button handed the whole thing off to his normal GitLab issue workflow. He still had to tell it to stop hallucinating a SAML login and eleven MCP servers that did not exist, but the result genuinely impressed him. Both hosts circled the same uneasy conclusion: for demos this clears the bar easily, and while serious applications still need real design and security judgment, these tools are coming for a lot of how people work day to day.

• Token Counting and “Token Maxing” (34:50)

  This turned into a confessional. Brett installed a token-counting status bar and discovered he had burned 167 million tokens in a day — climbing to 170 million by the end of the recording — and asked Travers to install the same tool to confirm whether he has a problem. The bigger number: the developer behind OpenClaw reportedly spent over a million dollars in tokens last month, running hundreds of bots across every repo with effectively unlimited OpenAI access. That fed into a discussion of “token maxing,” where tech companies measure engineers by token consumption as a proxy for activity. The hosts were not sold — monitoring metrics is fine, but token count is a Cobra-effect metric begging to be gamed.

• The $30K Claude-on-Bedrock Bill (38:30)

  The Register covered an AWS customer who ran up a $30,000 bill in April on Claude via Bedrock. What makes it sting is that the customer did the right thing — they had Cost Anomaly Detection configured — but anomaly detection does not cover AWS Marketplace, and Claude on Bedrock bills through a Marketplace subscription. AWS’s response amounted to “it’s in the documentation.” They had also burned through nearly $10K in credits, which Cost Explorer hides by default unless you know to remove credits from the view. Brett’s standing advice holds: set budgets, every account, no matter what you are doing — and his own takeaway was to go double-check the budget on his OpenBrain account, which also runs Claude through Bedrock.

• The Yarbo Robot Lawnmower Hack (43:09)

  Travers walked through the Yarbo robot lawnmower hack, since the original Verge piece is paywalled. A researcher found that Yarbo units are essentially unprotected and listening on the internet, meaning any and all of them can be hacked. To prove it, he hacked a journalist’s mower and ran the journalist over with it — blades off, thankfully. The hosts had grim fun with the implications: a hacked fleet of bladed lawn robots, or a food-delivery business built entirely on hijacked machines. The serious point underneath is real — regulation on security for these devices is coming, because semi-smart robots are about to be everywhere, and the Yarbo case sits right next to a recent humanoid robot that was fully controllable over Bluetooth.

• Recovering a $400K Bitcoin Wallet with Claude (46:48)

  The Register also ran the story of a man who recovered a Bitcoin wallet holding roughly $400,000 with Claude’s help. He had lost one of the passwords guarding it, so he fed Claude everything he could find — old notebooks photographed at his parents’ house, old work laptops — until it worked. Brett’s takeaway is the obvious one he repeats most weeks: use a password vault. The detail that got both hosts was the origin story, where the guy bought five bitcoins for a few hundred dollars at a Starbucks in 2015. The article apparently reveals his actual password, which Brett promises is worth the read.

• 1Password’s Agentic Monorepo Refactor (49:56)

  1Password published what they learned using AI agents to refactor a multi-million-line Go monorepo. Generating code was not the hard part — sequencing coordinated changes was: schema read/write ordering, shared state, deployment order, the wide-sweeping context across a big repo. They got a 20-30% lift on complex service extraction, and the monorepo structure helped because everything the agent needed was in one place rather than scattered across microservices. The lessons the hosts latched onto: have agents build their own deterministic tools, invest heavily in detailed playbooks so agents execute migrations exactly the way you want, and remember that without enough context an agent will confidently fill the gaps. Brett tied it to his own shift — instead of asking an agent to evaluate code against a standard, which was hit-and-miss run to run, have it write deterministic scripts that run in every pipeline. His prediction: our whole world becomes scripts and Markdown files within six months.

• Bedrock Prompt Optimization and LLM-as-a-Judge (57:09)

  Amazon Bedrock added an advanced prompt optimization and migration tool. It optimizes a prompt for any model on Bedrock and compares results across up to five models at once, iterating toward the best output. It is multimodal — PNG, JPEG, and PDF inputs — and you can guide the optimization with a Lambda function, an LLM-as-a-judge rubric, or a plain natural-language description. That introduced Brett to the term LLM-as-a-judge: prompting one model to assess, score, or give feedback on another system’s output. He drew a line to Claude’s skill-metadata optimizer, which test-runs a skill’s description to make it more likely to trigger when expected — the same loop-and-assess idea. Neither host expects to use the Bedrock tool soon, but for anyone optimizing across hundreds of thousands of requests, the cost and quality tradeoffs are worth the look.

• Cerebras IPO (1:02:01)

  Cerebras — the chip company whose wafers are bigger than your head — finally went public, raising $5.5 billion with the stock popping 108% in 2026’s first big tech IPO. Brett’s reaction was mock non-buyer’s remorse: retail investors could not have bought in anyway, and the hosts have been talking about a Cerebras IPO for so long it became a running joke. They noted the open question that sank an earlier attempt — customer concentration, with a large share of revenue tied to a single Gulf-region source. The segment dissolved into missing the Nvidia boat and waiting for the next one.

Off-the-Clock Recommendations

• They Will Kill You (2025) — Travers’s pick. A cult-in-a-hotel thriller he described as Kill Bill in a hotel. Entertaining, not too heavy, a good time.
• Send Help (2025) — Brett’s pick. Worth a watch, good for a couple of chuckles.
• Mad God (2021) — Travers’s pick, with caveats. Phil Tippett’s 30-year stop-motion passion project; very strange and genuinely gross imagery, but striking craft.
• Maximum Overdrive (1986) — Brett’s pick, prompted by the robot lawnmower story. Stephen King’s machines-turn-on-humanity film, complete with a man chewed up by a lawnmower.
• Lay’s World Cup chip flavors — Bacon Poutine, Steak and Chimichurri, and Cheese and Onion. Spotted ahead of the World Cup; expect a few bags around the Gillett house.