AWS Let Agents Write an Emulator With No Human Review
Agents have none of the slowness or fear that kept decades of unused permissions dormant, so they find your permission debt in minutes. AWS also shipped a local emulator covering 438 API models where agents wrote and reviewed every line.
Top-Line Summary
Monica Colangelo’s “accidental human firewall” is the idea I’ll be stealing for every talk from now on: we’ve relied on humans being slow, cautious and too busy to go looking for over-provisioned credentials, and agents have none of those traits. We also dig into AWS Emulator, a local emulator of 438 AWS API models written in Rust by agents with no human-written or human-reviewed code, and how Travers and I actually scope credentials for the agents we run.
Show Video
The “Pre-Show” Context
LinkedIn threw an error at me right as we started because I’d originally scheduled the stream for the 25th instead of the 18th (wishing November away, apparently) and it didn’t like the reschedule. re:Invent is about a week out and Travers flies in on the Sunday, which is the right call. I flew in on Mondays for years and it never worked.
The Engineering Rundown
-
AgentCore Updates and Agent-Driven Account Creation (01:38)
I didn’t read the AWS What’s New feed again this week, so Travers carried this one. AgentCore got a batch of updates: changes to Memory, free web queries, OAuth 2 support in Identity, and six more Regions for Runtime (from his notes on air; check the announcements for the details). You can also now start the AWS account creation flow from an agent. I’d seen the new developer path for creating accounts, and someone I talked to today called it a nightmare. If you already have an AWS Organization, use Organizations. I wrote tomorrow’s Bedrock Knowledge Base demo instructions against the new flow because it doesn’t need a credit card, so I’d better test it in the morning to make sure Bedrock is actually available.
-
AWS Emulator, Built Entirely by Agents (04:45)
AWS released a free local emulator of AWS services, and the article says agents wrote every line with no human review. It covers 438 AWS API models, it’s written in Rust, and it runs as a single binary on Linux or in Docker on macOS and Windows (per the announcement, which also notes it doesn’t enforce IAM policies and doesn’t take pull requests). Travers counted about 30,000 commits in three weeks. The build used agent teams: Test Writers check scenarios against real AWS, Issue Finders replay them against the emulator, and Fixers take one issue each through a merge gate. Travers liked the idea of allocating a pool of agents to wherever throughput is stuck. I used LocalStack for local testing on the Kanban project until it started demanding a licence, ripped it out for mocks, and haven’t missed it. This could be the replacement. It’s AWS writing about its own product, so I’ll try it on the train before I believe the fidelity claims.
-
Kiro Crew Bots (08:16)
Bots in Kiro Crew are named, long-running agents with their own memory, chat, notes and dashboard, and an ongoing responsibility instead of a throwaway session. The examples are a build-health bot that checks CI every 30 minutes and an issue bot that triages open issues. It looks a lot like something I built with Kiro Crew’s issue tracker a while ago: three agents pointed at a project, one for bugs, one for chores, one for docs, checking in the background. I tried to demo it live. Kiro Crew patched itself twice on air, my older sessions were blank, and the issue tracker was gone, possibly because I dropped from Kiro Pro+ to the free tier (I don’t know that for sure). Travers called these “smart cron jobs,” and I agree. Half the agents on my Fargate setup wake up every two hours, do one thing and go away. It’s fancy cron with a much better prediction engine behind it.
-
Fable 5.2 Burns Subscriptions (16:19)
Travers burned through both his subscriptions in about a day, mostly on Fable 5.2 orchestration and reviewing large codebases, then tried GPT Astra for the same work, didn’t love the results, and burned that too. His estimate is that Fable runs about three to one against Opus 5.5 in usage, roughly $3,000 of equivalent API spend per week versus about $7,000 if you stay on Opus (his numbers, unverified). He’s on the 20x plan and is considering a second one. He still logged 940 commits across 16 repos.
-
Sub-Agents Inherit Your Whole MCP Stack (18:06)
I hit my five-hour limit twice last week. The first was the skill benchmark from last episode, which fired up whole Claude instances with
claude -p. The second came while building sub-agents, even though only the planner runs on Opus and the rest run on Sonnet. What I hadn’t realised: sub-agents inherit the main agent’s environment, so every one of them loads all my connectors and tools, including Chrome and Google Drive, which they never use. I’ve started turning those off. Travers’ read is that headless sessions carrying your full MCP context will do exactly that. -
3D Modelling With Blender and FreeCAD (21:38)
Travers is designing a cyberdeck, a 3D-printed Raspberry Pi computer. He’s not a CAD person, but Astra and Fable 5.2 are good enough at Blender now to produce solid renders, which he uses to try different looks before moving to FreeCAD to design the actual parts. That surprised me less than it would have a month ago. Opus one-shotted the STL files for my agent status light after looking up the board dimensions, and the only defects are from my printer.
-
The Wall Dash (23:59)
I’ve got a 47-inch Toshiba Regza mounted up near the ceiling of my office. It’s so heavy it will never move, because I’m not getting crushed by a TV in a freak accident. I’m building a dashboard for it on a Raspberry Pi: time cards, what the agents are working on, what herdr is doing, all from the MCP servers I’ve replaced my SaaS tools with. Claude modelled the whole thing from a list of MCPs and it runs on my laptop. The brand-new Pi’s Ethernet port won’t light up, and I’ve tried a few cables. I also picked up a three-to-four-inch LCD with an Adafruit controller that might become version two of my agent status light.
-
Productivity FOMO (28:24)
Nobody serious measures productivity by commits, but watching talks where someone did 2,000 PRs still makes me wonder what I’m doing wrong. Travers’ take: that throughput needs a lot of supporting architecture before you can trust it in prod, and if you don’t need the volume, you can do a lot with a small environment you tend and tune over time. He calls it productivity FOMO. The real question is the value you get per dollar. Some of those people have essentially unlimited budgets, and we have one or two subscriptions.
-
Splitting Up My Go Skill (34:04)
My
/goskill used to grab one issue every time I typed “go.” Now there’s a separate plan skill: I describe a goal, Opus plans it, a script gates the plan, and it retries up to three times before stopping to ask me what it couldn’t work out, a mini grill-me. That loop is probably where a lot of my tokens go. It creates GitLab issues, labels each one (does a human need to look, does it touch cost, can an agent just do it), and runs agent-ready issues three at a time in sub-agents, which run the wrap skill when they finish. I test it all against a dummy “go sandbox” project in GitLab. The part I’m not happy with: it files Kanban cards for things that are really one question, like “Secrets Manager or Parameter Store?” A card should be something that takes me real time. I’m still working out the trigger. Travers thinks I’m moving the right way: I shouldn’t be the trigger for every task. -
Agent Credentials and Blast Radius (40:31)
I’ve been adding hooks so the agent always runs the evaluation script. For secrets, each project gets its own 1Password vault and service account, so an agent can only see its own credentials. That’s the right blast radius, but it doesn’t scale: I have a lot of vaults holding one GitLab token each. I’m surprised password managers don’t have better least-privilege tooling for agents yet, unless it’s in an enterprise tier. Travers runs any agent with AWS access in its own Docker container with a scoped role. Mine have IAM users in one central account that assume roles into their project accounts, and every access key carries a condition that makes it invalid after 60 days. A script Claude wrote now provisions the user, role and vault. No agent has more than read access. Deploys run through GitLab runners. Read-only is still a huge win: Claude builds CloudWatch filters and Athena queries over WAF logs in ten minutes that would have taken me two days.
-
The Accidental Human Firewall (49:39)
Monica Colangelo, an AWS Hero, argues that security has quietly relied on humans being slow, afraid of consequences, and too busy to go looking for over-provisioned credentials. Agents have a goal and no fear, so they surface decades of “permission debt” in very little time. Travers is waiting for an agent at work to find an old over-permissioned permission set, and suggests counting your access keys older than a thousand days. My twist: Security Hub has found these things for years, but a human still has to chase the owner three times. Could a trusted maintenance agent close that gap? Travers thinks auto-remediation has been the holy grail for a long time, but you still need guardrails. You don’t auto-rotate a thousand-day-old key that was used an hour ago. You find who owns it. A human is still accountable for risky actions. Her controls are all deterministic and AWS-native, and they apply to any environment.
-
Model Releases This Week (55:58)
Travers had an agent pull the week’s releases from Artificial Analysis: Gemini Decision 1 (yes/no answers), Gemini 4 Hydrogen, Qwen 4 Max, Tencent’s Hy4 Flash, GPT-6.2 Astra and Sol, and a batch of Microsoft MAI models. He also mentioned Google’s new frontier model, Neon. Google seems to be working through the periodic table. I use a small, cheap model on Cloudflare to tie Kanban cards to projects for the nightly brain summaries, and it works fine, but I may try Gemini Decision for it. On air I guessed it’s about 10 cents per million input tokens with free output. Check before you rely on that.
-
Agents Are Just Markdown Files (1:00:48)
Mathew Mathew argues most teams overbuild their agents. All you need is a folder of Markdown files for instructions, context, tools, invocation and feedback, and he walks through six questions for designing one. There’s a pitch for his open-source tool, Fife, at the end. I’ve overcomplicated plenty myself lately. Travers says Opus 5.5 needs far less scaffolding than models did even four months ago, and I remembered Steve Yegge saying something similar about why he stepped back from Gas Town (my recollection).
-
Cloudflare Kitesurf (1:02:47)
Kitesurf is a headless browser for agents that runs in V8 isolates on Cloudflare Workers, with no Chromium underneath. Travers says OpenCode integrated it almost immediately, and he plans to try it in his research pipeline. My pod bot uses Exa for web research, and its $10 of free monthly credits means it costs me nothing, so Kitesurf would have to earn the swap. It’s free in beta, and it can’t play video, render WebGL or get past bot challenges.
Off-the-Clock Recommendations
- Godzilla Minus One (2023): Travers’ anti-recommendation. He didn’t enjoy it, though a big Godzilla fan might.
- Ethan Mollick’s second book: I’m about 40% through and so far it covers a lot of the same ground as the first. I may leave it half done.
Claude drafts these notes from the transcript and my outline, and I review and edit them before they publish. Same rule I apply to any agent's pull request: it proposes, a human owns it.