Photoshop in Rust, Agent Inboxes, and a Text-Message Gym Buddy

Top-Line Summary

It’s pre:Invent season, and AWS is shipping agent blueprints faster than anyone is shipping agents: best practices for DevOps Agent skills, an AgentCore assistant built on OpenClaw, and Pizza Bot, an open source inbox for background agents. Brett and Travers ask where the production use cases are, then get into Mistral Large 4, Gemini 4 Argon, OpenAI’s math repo and a clean-room Photoshop in Rust. Brett also shows the two agents he built this week: one that turns his recorded calls into kanban cards, and one that texts him every afternoon to ask if he did what he said he would.

Show Video

The “Pre-Show” Context

Brett opens by admitting he’s feeling burnt out from trying to keep up with the churn. Travers is wondering whether all the harnesses he keeps building are worth it, or whether he’s better off sticking with the defaults and letting OpenAI and Anthropic figure it out, which is roughly where Steve Yegge landed with Gas Town once Opus made it unnecessary. Brett’s office lights misbehave for most of the first segment, and a thunderstorm rolls in during the second half.

The Engineering Rundown

• Best practices for writing AWS DevOps Agent skills (02:19)

  Brett still hasn’t tried DevOps Agent, but he thinks the post is good skill-writing advice for any harness: the SKILL.md format is the familiar name, description and instructions, and the description is still what decides whether a skill ever loads. The tip that hit home was “write investigation steps, not facts.” He admits he puts project facts into his CLAUDE.md files and skills that go stale quickly, when the agent can read the repo itself. He also likes the post’s idea of a last-verified date, because a skill-generating skill has left him with skills he suspects were used once. Travers wants a review step in the workflow, and Brett wants an analytics dashboard for skills usage, which they agree is one more project for a pair of people who just said they’re burnt out.

• A context-aware AI assistant on AgentCore and OpenClaw (07:59)

  The sample is a Telegram bot called Sprout running on AgentCore runtime and AgentCore Memory, and both hosts see the same thing every context engine now looks like: a wiki or retrieval layer, a bash tool and some context enhancers. AWS puts light use at $5 to $9 a month against about $35 for an always-on EC2 box, which Brett notes you could say about almost anything compared with EC2. His own context system has run on AWS since March for about $50 a month, and $30 to $35 of that is Aurora Postgres, even when it scales to zero. Travers paused his DynamoDB migration after finding PageRank works better than RAG for navigating his wiki. Their advice if you try the sample: set a budget alert that tells you about $10 in three days, not $10 in a month. Brett thinks OpenClaw already feels dated, which led to Jev: Travers got a key and found it a fine classifier, and Brett is still waiting for his.

• Pizza Bot, an open source inbox for background agents (13:37)

  Travers walks through the pattern: agents run tasks in the background and send their results to an inbox, which is turning into the standard way to manage long-running agents. Brett’s cynical read is that AWS publishes these because the production AgentCore stories aren’t out there, and if there were lots of them we’d see the blog posts and customer stories. Travers notes that a lot of people do the same thing with Slack or Discord threads. Brett wants it a step further: years of 2 AM disk-full pages have convinced him an agent should run the twelve-step runbook first and only wake him if it fails, and that some of it doesn’t need an agent at all, because Systems Manager automation documents already exist. He compares these samples to the espresso robot at every Summit, and predicts that at re:Invent, Bedrock and AgentCore get squashed back into one thing, since customers already ask him what the difference is.

• GLM 5.3 on Amazon Bedrock (21:33)

  Z.ai’s GLM 5.3 is another big open-weight model on Bedrock: 753B total parameters, about 40B active, a 1M token context. The specs barely registered. What both hosts like is not needing another vendor key, since Brett has keys scattered across Z.ai, Kimi and a Google account he’s not sure works. The catch for Canadian customers: it’s available through US and Global cross-Region inference only, so requests are processed in the US. Brett says he hasn’t found any of these models with a Canadian endpoint.

• Amazon Nova 2.5 Sonic (23:39)

  Brett’s reaction was “huh?”, because he thought Nova was being wound down. Travers figures it’s for Amazon Connect and might be the Nova model that survives. That led to Brett’s real voice news: he upgraded his Echo to Alexa+ through Prime and lost the calendar sync that used to warn him ten minutes before every meeting. He read somewhere that a certificate between the device and Google expired, which he hasn’t confirmed. The new voice was too energetic for the house, and he told the story of his first Echo, brought back from the US, which only worked once he gave it a postal code just over the border in Niagara. Neither host has a use case for speech-to-speech models, and Travers finds the ones that keep coming up dystopian: call centres and AI job interviews.

• Amazon Q Developer visualizations in the console (31:16)

  Second “huh?” of the night: Q Developer still exists in the console, with new health dashboards, charts, anomaly detection and troubleshooting. The post switches between “Amazon Q Developer” and “Amazon Q” and never mentions Kiro, which Brett thought replaced it. His other complaint is that this is one of 47 AWS services doing the same thing, since DevOps Agent already covers troubleshooting. Brett has only ever opened the console Q window by accident, and Travers tried the CLI a few times before switching to Claude.

• re:Invent plans (33:30)

  Travers is going to re:Invent for the first time and got into a couple of sessions before they sold out 30 minutes after opening. Brett’s advice is to watch the app for repeat sessions and new sessions added after keynote announcements. Brett is staying home again, partly because of a Toronto-to-Vegas return flight delayed by “a little damage to the plane.” They floated a simulcast from Vegas on the Wednesday night.

• Travers’s projects: LiDAR renders on GPU instances, and Melee recompiled (37:01)

  Travers is helping his uncle process drone LiDAR surveys, which are too heavy for a laptop, on AWS GPU instances. The hardest part was getting capacity at all, because GPU supply is that constrained. He built a computer-use skill for Codex on Linux with SSM-based screenshots and remote control, so Codex tunnels to the instance and clicks through the whole project setup. He only opened the GUI three or four times, to solve CAPTCHAs Codex refused to do. His Super Smash Bros. Melee decompilation and recompilation now builds for Windows with netplay mostly working, and he says decompiling old games has gotten easy enough that someone has Halo running in the browser faster than on the original Xbox.

• Brett’s call-capture agent (41:23)

  Brett records his calls with Granola and had a skill that pulled out everything he’d committed to, but he kept forgetting to run it between meetings. Now a Fargate task in his existing kanban environment starts every two hours, pulls new transcripts from Granola’s API, records each meeting ID in a DynamoDB ledger so it never processes one twice, writes a summary to OrbitBRAIN, and files kanban cards for anything he promised. It posts to Slack when it’s done, and he hasn’t missed notes on a meeting since. API keys live in Secrets Manager and are injected as environment variables only while the container runs. He used his own system over Granola’s Claude connector because his customer and project context already lives there. Next, he wants a classifier that sends pure research cards straight to his dispatcher.

• The Coordinator: an accountability agent over SMS (47:17)

  Brett pushed back on Jeff’s line from last week that the agents work for you, and built an accountability “gym buddy.” On Monday he sets the week’s commitments, and every weekday afternoon it texts him about the top one and expects an answer, marking it slipped if he doesn’t reply. When he started, Claude suggested Amazon Pinpoint, which is reaching end of support, so it now runs on AWS End User Messaging, which Brett says uses the same underlying API calls under a new name. The weak spot is the reply parser: there’s no working model in the loop, so anything outside its exact keywords gets rejected, even “I’m waiting on Travers, I’ll follow up Thursday.” Brett plans to add something cheap like Haiku, and Travers points out that’s exactly a Jev use case.

• Cloudflare Clef decision models (50:36)

  Clef is Cloudflare’s answer to Jev, API-compatible and with weights published under Apache 2.0. Brett signed up and got a key right away, after never receiving one for Jev. The headline number is latency: Clef-flash at a 38.8 ms median against 524 ms for Jev. Brett’s question is whether a classifier needs to be that fast if the work is asynchronous. Travers guesses Cloudflare gets the speed by running it in its edge data centres, which the post doesn’t confirm.

• Gemini 4 Argon (52:53)

  Google announced Argon, but it’s only available to trusted cyber defenders for now, the same early-access pattern other labs have used for security work. It tops Google’s own benchmark tables, and Brett questions how much benchmarks mean now that every model is trained to be good at them. The plan when it opens up: find a use case and kick it. Brett’s other hurdle is the Google Cloud console, which he admits is probably mostly user error.

• Mistral Large 4 (54:18)

  Mistral Large 4 is about 1 trillion parameters with 52B active, multimodal, trained from scratch in Mistral’s own European data centres, with weights promised by the end of the month. At $1.36 in and $4.18 out per million tokens, it’s about a quarter of Opus 5.5’s $4 and $20. Travers’s point: it’s the only top-ranked open-weight model from outside China right now, and for security work you want to run locally, he thinks it’s the best option. Its preview is live, unlike Argon’s.

• OpenAI’s math repo (56:18)

  Travers explains that both big labs are working on frontier math problems, and OpenAI has now posted results from an unreleased internal model, at about three hours of ChatGPT Pro thinking per result. On air he described about 500 problems with 93 solutions. The repo itself lists 722 manuscripts in 372 families from roughly 4,000 problems posed, never uses the word “solved”, and says some unformalized results “could have issues.” Brett noted the jump from high-school math to open research in a short time, alongside Anthropic’s wet-lab biology result, and Travers mentioned Anthropic has also posted non-trivial physics work. Brett is still annoyed he can’t buy a reasonably priced GPU for his gaming rig, though he concedes they’re doing more important things.

• PhotoCraft, Photoshop rebuilt in Rust (59:12)

  Travers described PhotoCraft as a clean-room rebuild of Photoshop in Rust, close enough for someone editing YouTube thumbnails. On air he said it was decompiled from Adobe’s binaries, but the README says it was built “from public specs and observed behaviour only” and calls itself early alpha. Brett’s prediction is a letter from a law firm with lots of last names, because Photoshop is Adobe’s cash cow: building something for yourself that never leaves your laptop is one thing, publishing it on GitHub is painting a target on yourself. Travers argues the clean-room process keeps the decompiling and the writing separate, and asks what Adobe can do when anyone with a $200 subscription can ask for their own photo editor. Brett’s reply: do this to CorelDRAW and nobody cares, but go after the biggest kid on the block and see what happens. Travers is waiting for someone to do SolidWorks.

Off-the-Clock Recommendations

• Evil Dead Burn (2026): Brett’s pick, and both hosts watched it. Extremely gory, and not the campy Evil Dead of the original cabin, which is why Brett found the Letterboxd reviews split between love and hate. Travers thought it was fine but “wasn’t Evil Dead to me,” and the dishwasher scene got him. Worth it for spooky season if you’re not expecting Evil Dead II.
• The ‘Burbs (1989): Travers’s pick. Early Tom Hanks: a strange new family moves in and the whole block decides they’re serial killers.
• Disturbia (2007): Brett’s companion pick to The ‘Burbs: Shia LaBeouf under house arrest, convinced his neighbour is a killer. Silly, with a great soundtrack.
• Obsession (2026): Brett figured out where it was going early, but rates the lead’s performance highly. If you’ve seen a wish movie before, you know how it ends.
• Backrooms (2026): Brett didn’t finish it, and his wife says he didn’t miss anything.
• Hokum (2026): Travers liked this Irish folk horror better than this year’s two big horror releases, and Brett would rewatch it before either of them.