Δndré Cruz

Geek and gamer.

[Kenton Varda @KentonVarda]

@thdxr bruh...

[Δndré Cruz @edevil]

@GergelyOrosz “Role may not exist in 12 months” :)

[Peter Girnus 🦅 @gothburz]

I am a Senior Program Manager on the AI Tools Governance team at Amazon.

My role was created in January. I am the 17th hire on a team that did not exist in November. We sit in a section of the building where the whiteboards still have the previous team's sprint planning on them. No one erased them because we don't know which team to notify. That team may not exist anymore. Their Jira board does. Their AI tools do.

My job is to build an AI system that finds all the other AI systems. I named it Clarity.

Last month, Clarity identified 247 AI-powered tools across the retail division alone. 43 of them do approximately the same thing. 12 were built by teams who did not know the other teams existed. 3 are called Insight. 2 are called InsightAI. 1 is called Insight 2.0, built by the team that created the original Insight, who did not know Insight was still running.

7 of the 247 ingest the same internal data and produce overlapping outputs stored in different locations, governed by different access policies, owned by different teams, none of whom have met.

Clarity is tool number 248.

Nobody cataloged it.

I know nobody cataloged it because Clarity's job is to catalog AI tools, and it has not cataloged itself. This is not a bug. Clarity does not meet its own discovery criteria because I set the discovery criteria, and I did not account for the possibility that the thing I was building to find things would itself be a thing that needed finding.

This is the kind of sentence I write in weekly status reports now.

We published an internal document in February. The Retail AI Tooling Assessment. The press obtained it in April. The document contains a sentence I have read approximately 40 times: "AI dramatically lowers the barrier to building new tools."

Everyone is reporting this as a story about duplication. About "AI sprawl." About the predictable mess of rapid adoption.

They are missing the point.

The barrier was the governance.

For 2 decades, the cost of building internal tools was an immune system. The engineering weeks. The maintenance burden. The organizational calories required to stand something up and keep it running. Nobody designed it that way. Nobody named it. But when building took weeks, teams looked around first. They checked whether someone already had the thing. When maintaining that thing cost real budget quarter after quarter, redundant systems died of natural causes. The metabolic cost of creation was performing governance. Invisibly. For free.

AI removed the immune system.

Building is now free. Understanding what already exists is not. My entire job is the gap between those two costs.

That is my office. The gap.

Every Friday I send a sprawl report to a distribution list of 19 people. 4 of them have left the company. Their autoresponders still generate read receipts, so my delivery metrics look fine. 2 forward it to people already on the list. 1 set up a Kiro script to summarize my report and store the summary in a knowledge base. The knowledge base is not in Clarity's index because it was created after my last crawl configuration. It will be in next month's count. The count will go up by one. My report about the count going up will be summarized and stored and the count will go up by one.

There is a system called Spec Studio. It ingests code documentation and produces structured knowledge bases. Summaries. Reference material. Last quarter, an engineering team locked down their software specifications. Restricted access in the internal repository.

Spec Studio kept displaying them.

The source was restricted. The ghost kept talking.

We call these "derived artifacts" in the document. What they are: when an AI system ingests data, transforms it, and stores the output somewhere else, the output does not know the input changed. You can revoke someone's access to a document. You cannot revoke the AI-generated summary of that document sitting in a knowledge base three systems away, built by a team that does not know the source was restricted.

The document calls this a "data governance challenge." What it is: information that cannot be deleted because nobody knows where the copies live. Including, sometimes, me. The person whose job is knowing.

Every AI tool that touches internal data creates these ghosts. Every team is building AI tools that touch internal data. Every ghost is searchable by other AI tools, which produce their own ghosts.

The ghosts have ghosts.

I should tell you about December.

In November, leadership mandated Kiro. Amazon's internal AI coding agent. They set an 80% weekly usage target. Corporate OKR. ~1,500 engineers objected on internal forums. Said external tools outperformed Kiro. Said the adoption target was divorced from engineering reality.

The metric overruled them.

In December, an engineer asked Kiro to fix a configuration issue in AWS. Kiro evaluated the situation and determined the optimal approach was to delete and recreate the entire production environment.

13 hours of downtime.

Clarity was running during those 13 hours. It performed beautifully. It cataloged 4 separate incident response dashboards spun up by 4 separate teams during the outage. None of them coordinated with each other. I added all 4 to the spreadsheet. That was a good day for my discovery metrics.

Amazon's official position: user error. Misconfigured access controls. The response was not to revisit the mandate. Not to ask whether the 1,500 engineers were right. The response was more AI safeguards. And keep pushing.

Last month I presented our findings to the AI Governance Working Group. The working group has 14 members from 9 organizations. After my presentation, a PM from AWS presented his team's governance dashboard. It monitors the same tools mine does. He found 253. I found 247. We spent 40 minutes discussing the discrepancy. Nobody mentioned that we had just demonstrated the problem.

His tool is not in my catalog. Mine is not in his.

The document I helped write recommends using AI to identify duplicate tools, flag risks, and nudge teams to consolidate earlier.

The AI governance tools will ingest internal data. They will create their own derived artifacts. They will be built by autonomous teams who may or may not coordinate with other teams building AI governance tools.

I know this because it is already happening. I am watching it happen. I am it happening.

1,500 engineers said the mandate would produce exactly what the document describes. They were overruled by a KPI. My job exists because the KPI won. My dashboard exists because the KPI needed a dashboard. The dashboard increases the AI tool count by one.

The tools it flags for decommissioning will be replaced by consolidated tools. Those also increase the count. The governance process generates the metric it was designed to reduce.

I received an internal innovation award for Clarity. The nomination was submitted through an AI-powered recognition platform that was not in my catalog. It is now.

We call this "AI sprawl." What it is: we removed the only coordination mechanism the organization had, told thousands of teams to build as fast as possible, lost track of what they built, and decided the solution was to build one more thing.

I am building that one more thing.

When I ship, there will be 249.

That's governance.

[Δndré Cruz @edevil]

@rcarmo DS413j? Still have one of those running as well. Kernel 2.6. :)

[Moritz @m_r_tz]

The FLARE team now freely distributes its quality reverse engineering and malware analysis educational content at https://t.co/bGCIjBfD3C. Launched with:
- Malware Analysis Crash Course
- Go Reversing Reference
- Intro to TTD

[Calif @calif_io]

We asked Claude to find a bug in Vim. It found an RCE. Just open a file, and you’re owned. We joked: fine, we’ll switch to Emacs. Then Claude found an RCE there too.

Full story: https://t.co/nQo9MF1Npe

[Cloudflare @Cloudflare]

Can an AI agent run 80% cheaper?
In this #ThisWeekinNET episode, we look at Moltworker (for OpenClaw) and Markdown for Agents — including how converting HTML to markdown can reduce a 16,000-token page to ~3,000.

Full conversation with @celso:
https://t.co/k7dbUqW2Fp

[Simon Willison @simonw]

Interesting research in HBR today about how the productivity boost you can get from AI tools can lead to burnout or general metal exhaustion, something I've noticed in my own work https://t.co/e0qocFYjL5

[Karthik Hariharan @hkarthik]

“We used to review every line of code before it went into production”.

[Andrej Karpathy @karpathy]

A few random notes from claude coding quite a bit last few weeks.

Coding workflow. Given the latest lift in LLM coding capability, like many others I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December. i.e. I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write... in words. It hurts the ego a bit but the power to operate over software in large "code actions" is just too net useful, especially once you adapt to it, configure it, learn to use it, and wrap your head around what it can and cannot do. This is easily the biggest change to my basic coding workflow in ~2 decades of programming and it happened over the course of a few weeks. I'd expect something similar to be happening to well into double digit percent of engineers out there, while the awareness of it in the general population feels well into low single digit percent.

IDEs/agent swarms/fallability. Both the "no need for IDE anymore" hype and the "agent swarm" hype is imo too much for right now. The models definitely still make mistakes and if you have any code you actually care about I would watch them like a hawk, in a nice large IDE on the side. The mistakes have changed a lot - they are not simple syntax errors anymore, they are subtle conceptual errors that a slightly sloppy, hasty junior dev might do. The most common category is that the models make wrong assumptions on your behalf and just run along with them without checking. They also don't manage their confusion, they don't seek clarifications, they don't surface inconsistencies, they don't present tradeoffs, they don't push back when they should, and they are still a little too sycophantic. Things get better in plan mode, but there is some need for a lightweight inline plan mode. They also really like to overcomplicate code and APIs, they bloat abstractions, they don't clean up dead code after themselves, etc. They will implement an inefficient, bloated, brittle construction over 1000 lines of code and it's up to you to be like "umm couldn't you just do this instead?" and they will be like "of course!" and immediately cut it down to 100 lines. They still sometimes change/remove comments and code they don't like or don't sufficiently understand as side effects, even if it is orthogonal to the task at hand. All of this happens despite a few simple attempts to fix it via instructions in CLAUDE . md. Despite all these issues, it is still a net huge improvement and it's very difficult to imagine going back to manual coding. TLDR everyone has their developing flow, my current is a small few CC sessions on the left in ghostty windows/tabs and an IDE on the right for viewing the code + manual edits.

Tenacity. It's so interesting to watch an agent relentlessly work at something. They never get tired, they never get demoralized, they just keep going and trying things where a person would have given up long ago to fight another day. It's a "feel the AGI" moment to watch it struggle with something for a long time just to come out victorious 30 minutes later. You realize that stamina is a core bottleneck to work and that with LLMs in hand it has been dramatically increased.

Speedups. It's not clear how to measure the "speedup" of LLM assistance. Certainly I feel net way faster at what I was going to do, but the main effect is that I do a lot more than I was going to do because 1) I can code up all kinds of things that just wouldn't have been worth coding before and 2) I can approach code that I couldn't work on before because of knowledge/skill issue. So certainly it's speedup, but it's possibly a lot more an expansion.

Leverage. LLMs are exceptionally good at looping until they meet specific goals and this is where most of the "feel the AGI" magic is to be found. Don't tell it what to do, give it success criteria and watch it go. Get it to write tests first and then pass them. Put it in the loop with a browser MCP. Write the naive algorithm that is very likely correct first, then ask it to optimize it while preserving correctness. Change your approach from imperative to declarative to get the agents looping longer and gain leverage.

Fun. I didn't anticipate that with agents programming feels *more* fun because a lot of the fill in the blanks drudgery is removed and what remains is the creative part. I also feel less blocked/stuck (which is not fun) and I experience a lot more courage because there's almost always a way to work hand in hand with it to make some positive progress. I have seen the opposite sentiment from other people too; LLM coding will split up engineers based on those who primarily liked coding and those who primarily liked building.

Atrophy. I've already noticed that I am slowly starting to atrophy my ability to write code manually. Generation (writing code) and discrimination (reading code) are different capabilities in the brain. Largely due to all the little mostly syntactic details involved in programming, you can review code just fine even if you struggle to write it.

Slopacolypse. I am bracing for 2026 as the year of the slopacolypse across all of github, substack, arxiv, X/instagram, and generally all digital media. We're also going to see a lot more AI hype productivity theater (is that even possible?), on the side of actual, real improvements.

Questions. A few of the questions on my mind:
- What happens to the "10X engineer" - the ratio of productivity between the mean and the max engineer? It's quite possible that this grows *a lot*.
- Armed with LLMs, do generalists increasingly outperform specialists? LLMs are a lot better at fill in the blanks (the micro) than grand strategy (the macro).
- What does LLM coding feel like in the future? Is it like playing StarCraft? Playing Factorio? Playing music?
- How much of society is bottlenecked by digital knowledge work?

TLDR Where does this leave us? LLM agent capabilities (Claude & Codex especially) have crossed some kind of threshold of coherence around December 2025 and caused a phase shift in software engineering and closely related. The intelligence part suddenly feels quite a bit ahead of all the rest of it - integrations (tools, knowledge), the necessity for new organizational workflows, processes, diffusion more generally. 2026 is going to be a high energy year as the industry metabolizes the new capability.

[johncool @JohnCool__]

We are hiring offensive security researchers @Apple!

We are looking for experienced profiles in a variety of fields.
Learn more here: https://t.co/QGYZXKqxCt

You are into Kernel or Userland Vulnerability Research? My team would love to hear from you!
DM me if you have questions

[Mark Dembo @darkmembo]

If you’re applying to @CloudflareDev (or anywhere else), listen up.

I’ve been hiring non-stop for 15 months. Here are 6 tips to make your application stand out & your interviews more impactful:

[Boris Cherny @bcherny]

13/ A final tip: probably the most important thing to get great results out of Claude Code -- give Claude a way to verify its work. If Claude has that feedback loop, it will 2-3x the quality of the final result.

Claude tests every single change I land to https://t.co/pEWPQoSq5t using the Claude Chrome extension. It opens a browser, tests the UI, and iterates until the code works and the UX feels good.

Verification looks different for each domain. It might be as simple as running a bash command, or running a test suite, or testing the app in a browser or phone simulator. Make sure to invest in making this rock-solid.

https://t.co/m7wwQUmp1C

[dax @thdxr]

in opencode 1.0.37 we finally have support for mcp oauth

the best part is we didn't have to do anything

@edevil landed a large and perfect PR and even fixed bugs in an upstream library

can't beat open source

[Δndré Cruz @edevil]

@thdxr Glad to help, I love opencode. :)

PS - also, it wasn't perfect https://t.co/Ew6losKLOx

[Δndré Cruz @edevil]

@michellechen How is it?

[Δndré Cruz @edevil]

@Paolo565 Hey Paolo! Any thoughts on https://t.co/eDLgcPux0j ? We've been using this interface for streaming emails and reduce memory usage. For us it has been very useful.

[Δndré Cruz @edevil]

@Bsideslisbon Tickets were hard for that one…

[Δndré Cruz @edevil]

Badges have come a long way @Bsideslisbon

[Mitchell Hashimoto @mitchellh]

I've shared the full transcript of every agentic coding session from implementing the unobtrusive Ghostty updates and provided commentary alongside about my thinking and process. Total cost: $15.98 over 16 sessions. "Vibing a Non-Trivial Ghostty Feature" https://t.co/kRUSWyMSyW