Deep dive · Xiaohu explains

AI Engineer World's Fair closes with a fight over whether coding-loop hype has outrun engineering discipline

The closing vote never got counted — the stage lights were too bright to see hands. A parallel survey found 95% of teams already use agents, and 59% worry they're racking up technical debt

Five-sentence recap

On the last day of the AI Engineer World's Fair (AIEWF), a live debate over the "loop" — an agent that writes code, runs tests, catches errors, and fixes them itself, on repeat — put the conference's underlying argument on the table: believers say there's no going back to hand-writing code; skeptics say the hype has outrun engineering discipline
Anthropic's Head of Labs and Instagram co-founder Mike Krieger talked about Claude Tag, the company's internal tool — the instruction given to it now is "you own this part of the codebase, watch the feedback channel yourself, and pick up work proactively," not fix bugs one ticket at a time
Amplify's 2026 AI engineer survey: 95% of respondents now use agents (up from about half last year), and 89% of those agents can write data (versus 52% last year); but 59% worry AI-generated code is piling up long-term technical debt, and the "control layer" for agents is still primitive
The closing keynotes turned optimistic: Theo Browne said "what used to be startup-scale is now just a side project"; YC president and CEO Garry Tan said the fastest-growing founders treat AI as a workforce, not autocomplete
The live vote fizzled out — the stage lights were so bright that neither the moderator nor the debaters could count how many hands went up

1 The live debate

On the last day, a fight breaks out over the "loop"

On the last day of the AI Engineer World's Fair (AIEWF), a live debate over the "loop" put the conference's underlying argument on the table. A loop is what happens when an agent writes code, runs tests, finds errors, and fixes them itself — round after round — without a human watching every step.

The debate set out to settle one thing: can a self-running "software factory" actually be deployed at scale today, or has the engineering discipline never caught up to the ambition? Moderator Allie Howe put it bluntly from the start: "Is there a gap between the hype around loops and how usable they actually are?"

📊

Why it matters: In Amplify's 2026 annual survey, 95% of surveyed engineering teams are already using agents — double last year's figure. In the same survey, 59% worry that AI-generated code is creating long-term technical debt. This debate lives right in that gap.

Arguing for was Geoffrey Huntley, creator of Ralph Loop, joined by Keycard CEO Ian Livingstone. Arguing against were Dex Horthy of HumanLayer and Greg Pstrucha of Subroutine. Neither side denied loops are useful — the fight was over whether coding can be handed over to agents wholesale, right now.

What a loop actually is

Hand a task list to an intern and let them try, fail, and retry on their own until it's done — instead of walking them through every step yourself. That's roughly what an agent's loop is: it runs on its own, and you're just there to make sure it doesn't go off the rails.

"We're basically the engineers making sure the train doesn't derail." — Geoffrey Huntley

The AI Engineer World's Fair floor. Source: Latent.Space

2 The case for

The case for: the loop is already here, and there's no going back

Huntley opened by saying the loop has already arrived. "This is inevitable — it's here to stay," he said, adding, "I don't think I'll ever go back to writing code by hand."

Livingstone picked up the thread and pushed the argument onto harder ground: what matters is verifiability. It doesn't matter how the code got written — by a human or an agent — as long as the result can be verified. He also pointed out that the loop has always been the core of software development; it's nothing new.

The pro case

In Livingstone's words: "The core of a loop is just: I try something, I learn something, I apply it. What we're really talking about is how fast we can spin that process." To him, agentic coding just spins this ancient loop faster.

3 The case against

The case against: the hype has outrun the discipline

Horthy opened by drawing a line: he's not against loops. "The basic question here isn't whether loops are good or bad," he said. Kubernetes itself is built on control loops, he noted — but those are deterministic loops. His real objection: "The hype has outrun the discipline."

What does "deterministic" mean here? The rules are fixed in advance — if server load spikes, spin up another instance; exactly how to respond is decided ahead of time, with no on-the-fly judgment call. An agent's loop is different: what happens at each step is a judgment the AI makes on the spot, so running the same task twice can produce a different process and a different result. The two kinds of loop look alike, but they're nothing alike underneath.

The agent's loop

The AI decides on the spot — run the same task twice and you may get different results

Kubernetes's deterministic loop

Rules are fixed in advance — same input, same result, every time

Horthy pushed the distinction further: what people call the "abstraction level" is really just how far removed a human is from the details — the higher the level, the fewer details a person manages, and the more gets handed to the AI. "I haven't seen evidence that we can move up an abstraction level right now," he said — meaning handing coding over wholesale to agents. "If anything, the direction should be to drop down a level."

Pstrucha was worried about a different ledger: money. He said the economic sustainability of agent loops is questionable — you can't "orchestrate your way out of a problem by buying more tokens."

PRO

Huntley / Livingstone

The loop is already here — no going back to hand-writing code
What matters is verifiability, not how the code got written
The loop has always been the core of software development: try, learn, apply

CON

Horthy / Pstrucha

Not against loops — Kubernetes has run on control loops for years, but those are deterministic
The hype has outrun engineering discipline
No evidence we can move up an abstraction level — if anything, we should drop one down
Not economically sustainable — you can't orchestrate your way out by buying more tokens

"Right now we're basically locomotive engineers. That's the job: keep the train on the tracks."Geoffrey Huntley, loop proponent

4 The software factory

The software factory's dilemma: once it's automated, who still touches the problem?

The conversation turned to the "software factory" — a metaphor, now popular across the industry, for handing the entire pipeline (writing code, testing, shipping) to a fleet of agents running on autopilot, with humans stepping back into an oversight and review role.

Horthy's worry was concrete: when everything runs automatically inside a factory-style agent environment, "you never actually touch the problem itself." So he suggested starting small — iterating with agent loops bit by bit to first "build intuition," rather than jumping straight to end-to-end automation.

Even Huntley admitted there's danger in the loop. He said the software factory represents where things are headed, but the industry hasn't solved it yet. "This is frontier thinking."

As the hour-long debate wrapped up, Howe asked the audience to vote by show of hands for who "won." What followed was a joke on humans themselves: the stage lights were so bright that neither she nor the debaters could tell how many hands were up. Maybe an agent should have been in charge of dimming the lights.

5 Anthropic's template

Anthropic's template: what Claude Tag actually looks like

If you're looking for a company actually moving toward the software-factory model, Anthropic is one. Instagram co-founder and current Anthropic Head of Labs Mike Krieger sat down for a morning conversation with swyx.

He talked about Claude Tag, the internal tool the company announced publicly the week before. He described Tag as more "delegated, asynchronous, and proactive" than Claude. This might be what an early-stage software factory actually looks like: agents haven't replaced the team — multiple people are delegating their own responsibilities to a system like Claude Tag.

▍ Additional source: Anthropic's official announcement, "Introducing Claude Tag"

Krieger only touched on it briefly on stage, but the official announcement spells Claude Tag out: it puts Claude into Slack as a team member. Give it access to designated channels, tools, and codebases, and anyone in the channel can @Claude to hand off work and go do something else; it retains relevant context from the channel over time, and it can even schedule future tasks. Anthropic positions it as the next evolution of Claude Code — a model that's more proactive and built for a whole team to use together.

Compared with opening a one-off chat window to task Claude, @Claude differs in four ways:

Multiplayer

There's just one Claude per channel, and everyone collaborates with the same instance. Anyone can see what it's doing and pick up where the last person left off — more like working with a colleague than running separate conversations.

Learns over time

It tracks the channel and accumulates context automatically, so nobody has to re-explain from scratch each time; with authorization, it can also learn automatically from other channels and data sources (private channels stay private), building up the tacit knowledge it needs to do the job well.

Takes initiative

With ambient mode turned on, it proactively surfaces information it thinks you should know, and stays on top of threads and tasks that have gone cold and unresolved — without anyone having to nudge it.

Asynchronous

Once you hand off a task, you can go do something else while it keeps working; it can even schedule itself, pushing a project forward for hours or even days at a stretch. Anthropic says a large chunk of their time now goes into delegating work to a fleet of Claudes in parallel.

One number, plus guardrails

According to the company, 65% of the code from Anthropic's product team is now generated by the internal version of Claude Tag (a vendor self-report), and the pattern is spreading beyond engineering into tracking product metrics, handling support tickets, and chasing down gnarly bugs. On the governance side, admins can carve out isolated "Claude identities" per channel — the memory and tools from the sales instance don't leak into the engineering one — and can also cap token spend and pull a complete log of who told @Claude to do what, and when. It's currently in beta for Claude Enterprise / Team customers, running on Opus 4.8, and it replaces the old Claude in Slack app.

How to delegate to Tag

"Most of our usage is really more like delegation," Krieger said. He gave an example: "Don't just fix this bug. From now on, you own this part of the codebase — I want you watching this feedback channel and picking up work proactively." He said this has already changed how the team operates, turning it into a "multiplayer, asynchronous, proactive" style of collaboration.

But he also pointed out a side effect of automation: the team is now "bottlenecked at review," and bottlenecked on "whether humans can actually hold the full picture of what we're doing in their heads."

Mike Krieger in conversation with swyx at AIEWF. Source: Latent.Space

6 The hard numbers

The numbers: the 2026 AI engineer survey

Back to the day-to-day reality of most AI engineers. That morning, Amplify's Barr Yaron released her annual industry survey.

According to Amplify's data, 95% of respondents now use agents — roughly double last year's figure. Among teams using agents, 89% say those agents can write data, up from 52% last year. "Agents aren't just reading, summarizing, and drafting anymore," Yaron said. "They're taking real action inside the system."

95%

of surveyed teams already use agents — roughly double last year

89%

of agent-using teams let their agents write data (up from 52% last year)

Using agents · 2025

~half

Using agents · 2026

95%

Can write data · 2025

52%

Can write data · 2026

89%

But the guardrails are still primitive. Human approval and permissions are the two main lines of defense, trailed by a scattering of other techniques: task decomposition, retrieval, memory, sandboxing. "Nobody has settled on a control layer for agents," Yaron said. In plain terms: agents can already take action inside real systems, but the industry still hasn't figured out how to rein them in.

Amplify's Barr Yaron presents the 2026 AI engineer survey. Source: Latent.Space

7 The central tension

Cheaper, and more anxious: cost and technical debt

Cost is another sore spot. 40% of respondents say AI costs "frequently" constrain their ambitions for using AI, and another 36% say it happens "sometimes." Token usage is now the second most-watched production metric, right behind quality.

Cost "frequently" limits ambition

40%

Cost "sometimes" limits it

36%

Worried about accruing debt

59%

The core tension

AI makes experimentation cheaper and lets teams ship more software, but among that same pool of respondents, 59% worry that the code AI generates today is building up long-term liabilities. The time and money saved may be getting quietly eaten back by the debt piling up behind it. That contrast makes the point more plainly than anything said on the debate stage.

8 Closing

The closing keynotes: what comes next

The conference's final sessions pulled the mood back toward optimism — looking at AI, and building things with it. That is, after all, the point of AIEWF, and the fun part too.

Theo Browne showed off a few projects he's built, or is still building, with AI. His point: the scale a single developer can realistically take on has shifted. "What used to be startup-scale is now just a side project," he said. Projects he'd have written off as "too big" in the past are now within reach.

YC president and CEO Garry Tan then brought that optimism down to the organizational level. He said YC's fastest-growing founders all treat AI as a workforce to deploy. Where most people used to treat it as glorified autocomplete — a couple of lines popping up while you type — this cohort hands it whole chunks of work.

YC president and CEO Garry Tan delivers the closing keynote at AIEWF. Source: Latent.Space

Back to that opening locomotive metaphor: a week's worth of debate made it clear how much engineering still stands between today and an AI-native vision that actually works for everyone. The closing keynotes were a reminder of why the engineers in the room keep chasing it anyway. In the words of the original piece, they just want to drive those locomotives.

"Build an AI-native company, not a company that just happens to use AI."Garry Tan, YC president and CEO, closing keynote

This piece is Xiaohu's take on Latent.Space's "AIEWF Daily Dispatch: The great loops debate and the state of AI engineering" (by Richard MacManus, July 3, 2026). All data, quotes, and people mentioned are drawn from the original article; survey figures are from Amplify's 2026 AI engineer annual survey.

One-Pager · read this ≈ read the whole piece

AI writing its own code: the argument shifted from 'can it work' to 'dare we hand over everything now'

A closing debate and an industry survey at the AI engineer conference — this one page, with charts, covers the whole argument over whether the hype has outrun engineering discipline.

↓ One page, and one chart that moves

1 First, the basics · what a loop is, and where it gets stuck

The closing session on the last day of the AI Engineer World's Fair was a debate over the "loop." A loop (an automated coding cycle) is when an agent — an AI program that acts and makes decisions on its own — writes code, runs tests, catches errors, and fixes them, round after round, with no human watching every step. Neither side disputed that it's useful; the fight was over whether it can be handed whole swaths of real work at scale, right now.

✔ It writes, tests, and fixes itself, cycling through until the job's done — saving a huge amount of manual labor
✘ But every step is a judgment call the AI makes on the spot — run the same task twice and both the process and the result can differ
Why it matters: when something breaks, humans "never actually touch the problem itself"; the industry still hasn't settled on how to rein it in (what insiders call the "control layer")

2 On the table · what the two sides are actually arguing

Arguing for: Ralph Loop creator Geoffrey Huntley and Keycard CEO Ian Livingstone. Arguing against: Dex Horthy of HumanLayer and Greg Pstrucha of Subroutine. Neither side disputes that loops are useful. The pro side says there's no going back to hand-written code — as long as the result can be verified, it doesn't matter who wrote it. The con side says they're not against loops; they're against the hype outrunning engineering discipline, and the economics don't add up.

PRO

The loop is already here — no going back to hand-writing code
What matters is whether the result can be verified, not who wrote the code
Companies are already doing this: Anthropic has its internal tool Claude Tag "own an entire part of the codebase, watch the channel, and pick up work on its own"

CON

Not against loops — against the hype outrunning engineering discipline
No evidence we can hand coding over wholesale right now — if anything, the approach should be more cautious
Not economically sustainable — you can't bury the problem under more tokens (the billing unit for how much text an AI processes)

The biggest disagreement really comes down to one word: which kind of "loop" are we even talking about?

3 One chart · the difference between the two kinds of loop

Horthy's key argument on the con side: loops have been around for a long time. Kubernetes (a widely used server-automation system) is built on loops too — but those are "deterministic," with rules fixed in advance. An agent's loop is the AI deciding on the spot. The two look alike but are nothing alike underneath.

The con side's line about "hype outrunning discipline" is exactly what's happening in the top row: an agent's loop can produce a different result every round, and it still runs on ever more compute — not yet mature enough to hand over an entire codebase. The deterministic loop in the bottom row is the proven approach the industry has actually used for years.

4 The numbers · where the industry actually stands

At the same event, Amplify's Barr Yaron released the annual survey. Translate the percentages into something concrete: of every 20 teams surveyed, 19 are already using agents — a year ago, that figure was only about half. But among that same group, more than half worry they're racking up technical debt (the shortcuts taken to move fast now that cost more time to pay off later).

95%

of surveyed teams already use agents — up from about half a year ago, nearly double

89%

of agent-using teams let it write data directly (up from 52% last year)

76%

say AI cost limits their ambitions (40% frequently + 36% sometimes)

59%

worry that AI-generated code today is piling up long-term technical debt

All of these figures come from Amplify's 2026 AI engineer annual survey, self-reported by the engineering teams surveyed, with no independent replication.

⚠️

This is the conference's central tension: AI makes experimentation cheaper and lets teams ship more, but the time it saves is being quietly eaten back by the technical debt more than half of respondents are worried about.

The one line to take away: nobody's arguing anymore about whether AI can write its own code — it's already doing that at 95% of teams. What's still unresolved is how to hand over whole chunks of work without turning today's time savings into tomorrow's technical debt. The direction the closing keynote pointed to: treat AI as a workforce, not the autocomplete that pops up a couple of lines while you type.

Sign in