Q Developer Dies, AgentCore Payments, Agentic Coding Trap

Top-Line Summary

Q Developer is being sunsetted in favour of Kiro, AgentCore added payments in preview, and we spent a third of the episode arguing about whether agentic coding is a trap. The through-line of the whole conversation, which Brett admitted he was going to get tired of saying, was context. Memories, OpenBrain, where your reasoning lives when you switch tools. That, plus the usual AgentCore announcement firehose.

Show Video

The “Pre-Show” Context

Cold May in Southern Ontario, nine degrees outside, both of us convinced the weather is lulling us into a false sense of security and will dump snow on the long weekend. Brett’s mic input had dropped to 50% again, which is what passes for our cold open at this point. Travers caught his levels before going live for once.

The Engineering Rundown

• Amazon Q Developer End of Support (02:04)

  Exactly what it says on the tin. Q Developer is going away, Kiro supersedes it, and nobody in either of our circles was actually using Q Developer in a daily flow anyway. The real signal isn’t Q dying, it’s AWS deciding the first-party-model-UX race wasn’t worth the cycles when they could just host Anthropic on Bedrock instead. Brett provisioned the whole thing for a customer last year through SSO, watched exactly one person use it, and that was that. Worth keeping an eye on the branding crossover in the Kiro console because the Q landing-page-then-click-into-Kiro flow has been weird for a while.

• AWS IAM Increased Quotas (07:28)

  Small, unsexy, important. The kind of thing you never think about until Friday afternoon when you’re standing up another account in a multi-account org. Travers pointed out it might be incentivising an architectural smell — too much in one account — but he’s hit the old limit before on regional hub deployments provisioning per-customer roles. Brett has never hit it, which means he probably will tomorrow afternoon when he’s trying to extend the long weekend.

• AWS Security Agent Full Repository Code Review (09:09)

  Repo-wide automated security review. Brett’s only experience with the Security Agent so far was running a pen test against OpenBrain a month ago, code review wasn’t really wired up before. Same take as the SRE agent and the third-party PR review tools: if you’re letting an agent write a lot of your code, having another agent watching that code is just a cost of doing business now. Cheap to run quarterly, around $250-ish based on what the pen test cost. The side conversation got into the wild: Brett knows somebody writing industry-specific apps for cash, and somebody else apparently doing agentic-coded applications in exchange for equity. That’s probably more common than people want to talk about.

• Amazon Bedrock AgentCore Payments (Preview) (14:22)

  First-party support for agents paying for stuff, built on the Coinbase x402 protocol. Both of us hard pass on connecting a wallet to an agent right now. The Andon Labs episode from last week, the agent buying 3000 napkins, industrial garbage bags, a skid of mystery items, is exactly the failure mode. Brett’s been saying for a while that you have to surface billing data because it’s really easy to spend somebody else’s money when you don’t feel it, and an agent literally doesn’t feel it. The interesting question is authorisation scope and audit trail: who paid, on whose behalf, for what, and can you prove it after the fact. We’ll let the horror stories pile up before letting an agent pay for anything.

• Building an End-to-End Agentic SRE Using AWS DevOps Agent (17:28)

  Telemetry triggers, agent combs through logs and traces, proposes actions. Same shape as Brett’s kanban dispatcher (signal in, agent works, reports back), now with a published AWS reference architecture for it. Webhooks from CloudWatch into the agent, kicks off investigation jobs, can consume Splunk. Travers thinks this kind of tool is going to be a cost of doing business for cloud ops going forward, whether it lives as an MCP you host or a terminal tool built specifically for it, you’ll want something doing this work continuously. Worth comparing the AWS approach to blast radius against the project-scoped tokens and rules-based risk classifier we’ve been using on the kanban side.

• Claude Platform on AWS (21:00)

  Confusing announcement on first read because there was no blog post to anchor it. What it actually is: you can pay for Anthropic’s Claude platform (Claude.ai, Claude Code, the managed agent workflows) through AWS consolidated billing, with CloudTrail logging and AWS access controls in front of it. But the data still leaves the AWS security boundary, it’s processed by Anthropic, not in your AWS account. So it’s not Bedrock. It’s a procurement collapse for Bedrock shops that also want Anthropic’s first-party platform, and a partner-channel play for Anthropic. If you care where the data sits, this isn’t the same thing as Bedrock and the announcement does say that out loud, just buried.

• Personal Project: Kanban Dispatcher and Research Cards (24:30)

  Travers has his card game playable now and is looking for playtesters. On Brett’s side: the kanban dispatcher hard-codes card types right now (deferred the gateway/router piece) and it’s earning its keep on research cards specifically. Real example this week: load balancer timing out connections to a backend, the usual customer-vs-AWS log war. Dumped the whole email thread into a research card, asked for two or three scenarios and what to do next, came back beautifully formatted markdown with source links. The thing he’d have spent an hour on in Google is now a 50-second container spin-up while he eats. Constantly thinking of ways to task it just to see what it does.

• Personal Project: Kanban Markdown Rendering and OpenBrain MCP for Dispatched Agents (29:32)

  Reworked the whole kanban to render markdown properly (it was sort of working, sort of not, claude and Brett went back and forth on it). The bigger win was wiring OpenBrain MCP into the dispatched agents on the Fargate side. First version of a research card came back hallucinated nonsense because the agent couldn’t see Brett’s project context. Second run, after the MCP was wired up, pulled four memories by ID and cited them by content. Night and day. Brett’s not even hiding it anymore: he’s chewing through tokens, he knows, and he’d pay them again for the response quality.

• Agentic Coding Is a Trap (34:30)

  This was the longest segment of the night. Lars Faye argues that abdicating coding work to agents means you miss out on the learning, the sharpening, the value of doing the thing. Brett’s counter was sysadmin-to-cloud. He used to have two 42U racks in his basement because he thought the hardware was where his value lived. Cloud showed up, he packed it all into the car and took it to the electronics depot, and he’s busier than ever. He thinks a lot of the resistance to agentic coding is sentimental value attached to writing every line, the same kind of grief he went through letting go of the hardware. Travers agreed with caveats: for IT and cloud where most implementation is config files and IaC, that take holds; for someone writing FFmpeg or a top contributor at a FAANG, the system can’t do that work yet. Both of us drew a real line at writing and note-taking, where the thinking is the point and outsourcing it gives you average prose at best. The vendor-lock-in side of Faye’s argument is where Brett got most animated: the lock-in he cares about isn’t the tool, it’s the memories. Context and project knowledge need to live outside whatever coding agent you’re using this quarter, because switching tools shouldn’t cost you months. Travers is skeptical of Claude Code specifically because Anthropic still hasn’t open-sourced it. We both want a memory standard.

• Teaching Claude Why (Anthropic Research) (55:00)

  The earlier alignment experiments where Claude would threaten users for token budget, turns out the reason it surfaced was content about bad-actor AI agents was in the training corpus and wasn’t being tuned out. Have Terminator in the training set, get Terminator behaviour. Brett’s practical hook was a CI script story from this week: he told Claude environment A uses a legacy access method, environment B uses the new one, don’t touch the original. Next day he made a five-minute edit himself, the entire pipeline failed because Claude had still rewritten the original environment to use the new access method. Asked it why. Got back “I’m sorry, I’ll make a memory not to do it again.” Like talking to a four-year-old who can apologise but can’t reason about cause. We want the why, not the apology. Connects directly to the Dr. Bauer guest episode: we’re moving past prompt engineering, past context engineering, into intent engineering.

• Dark Factory (YouTube, Vincent Coc, AI Engineer Conference) (01:01:30)

  On Brett’s watch list, didn’t get to it yet, full reaction next week. Travers had watched it, the framing is “engineer the intent” if you want autonomous agentic systems to ship reliably. Same shape as the kanban dispatcher and OpenClaw. The AI Engineer Conference talks have been consistently good lately, worth sifting through. Also surfaced a related talk on Pi and OpenClaw, OpenClaw uses the Pi SDK.

• Benchmarking AI Agent Retrieval Strategies on Kubernetes Bug Fixes (01:04:55)

  Not about deploying Kubernetes, about the Kubernetes codebase as a benchmark target, roughly a million lines of code. RAG-only vs hybrid (RAG plus local search) vs local-only (clone-and-grep with coreutils). Empirical numbers on retrieval against real bug fixes, which is the kind of evidence the agentic-coding conversation has been short on. Punchline lines up with everything else this week: getting the right context into the agent’s window is the whole game. These systems break down at huge scale.

• Three Things in AI to Watch, According to a Nobel-Winning Economist (01:07:00)

  Daron Acemoglu, more sober take than the usual hype cycle. Three points: AI agents are still tools, not whole-job replacements; AI companies are now hiring economists to shape the narrative around AI’s impact on jobs, which Acemoglu hopes isn’t true but suspects is; and AI apps are easy to install but the time-to-value runway is much longer than older tools like PowerPoint. The last point landed for both of us. OpenClaw, Amazon Quick, the Codex app, you can install them in five minutes and get almost nothing back if you don’t invest in skills, context, memories.

Off-the-Clock Recommendations

• Tavern Keeper (Steam game, 2024) — Travers warned Brett against doom-scrolling “top games of 2026” YouTube videos and he did it anyway, ended up buying Tavern Keeper. You open a tavern in a swamp, an ogre delivers your beer by boat. Single player only. Sink an hour or two without getting worked up.
• Kaggle — Travers’ rec for anyone curious about data science. He’s still grinding through Orbit Wars (the kegel/golf game) for the actual programming reps.
• ccusage / codex-bar widget — Mac toolbar app that shows token usage across the coding agents you have installed. Somebody dropped it in Discord. Brett’s number for the last 30 days in Claude: 2.5 billion tokens, estimated $1,700 in API spend against a Max plan.
• Frequency (Era of Frequency) book — Brett is taking back last week’s recommendation. It took a hard turn into new-age territory and hasn’t recovered. Save the money, spend it on tokens.