AI Builders Digest

Claude has been maintaining Anthropic's apps: 388 PRs opened, 180 merged

For the past few weeks Claude has run daily maintenance routines across iOS, Android, Desktop, web, CLI, and the Agent SDK, coordinated from a Slack channel. The routines include a crash fuzzer that taps around a simulator to find crashes and then root-causes them, a duplicate unifier that files PRs merging slightly divergent abstractions, and a dead-code remover that adds logging to suspected dead code and deletes it the next day if nothing fires. Of 388 PRs opened, 180 got merged after Claude Code Review plus human review. When a routine gets something wrong, the fix is asking Claude to tune the routine itself, which sometimes takes a few days of iteration.

• #agents
• #products

Gemini 3.7 Flash lands 50% cheaper, built in about three weeks

The new Flash model is faster, half the price of its predecessor, and went from start to ship in roughly three weeks. The turnaround time is the real signal here about how fast the release cadence has compressed.

• #products

ARC-AGI-3 nearly falls to a coding harness, and Masad calls it vindication

Adding a coding harness gets a model close to solving ARC-AGI-3, which Masad reads as confirmation of a thesis he has been pushing: coding is the capability that generalizes LLMs to everything else. He paired it with a sharper prediction, that by next year using a computer will be optional and work changes radically.

• #agents
• #evals

The engineer-elimination thesis was "absurdly wrong," and expertise gets more valuable

What actually happened is that engineers got a power tool, and their value goes up rather than down because engineering can now be applied to far more work than before. Automating drug discovery needs engineers. Automating manufacturing needs engineers. Taking on larger software projects needs engineers. The generalization: AI causes companies to take on more work, which means more experts to oversee it, and better models get better utilization from experts than from novices.

• #agents
• #products

Prompt debt is the new tech debt: cut half your prompt with every model update

Most AI products carry bloated system prompts that actively hurt quality. The accumulation pattern is familiar: model fails at something, add ten rules; tool call fails, add ten examples; output looks weird, add formatting constraints. Three months later the system prompt is a novel, and meanwhile the underlying model got much smarter while your micro-managing rules turn it back into a dumb rules machine. The prescription is aggressive: delete at least 50% of your prompts on every model update. He also counted more than 20 AI products named some variant of "Studio."

• #agents
• #products

Rauch bets one token-routing command becomes the default way to run coding AI

The pitch is a single command that configures every coding harness you already use, Claude Code and Codex included, rather than replacing them. What it buys you is uptime, model choice, lower costs, observability, and zero data retention. He is predicting this becomes the default way coding AI gets used at scale, and is currently offering free GLM 5.2 at up to 500 tokens per second through it.

• #products
• #agents

Chess is bigger than ever thirty years after computers beat humans at it

Chess.com started as a $56,000 domain bought out of a 2005 bankruptcy auction after every investor said it was uninvestable, and now runs 10 million daily actives, 250 million registered members, and over $200 million in annual revenue with no capital ever raised for growth. The counterintuitive part is what engines did to the game: Stockfish made chess boring for a while as players ground out perfect end games, then self-taught neural nets like Leela started beating it with aggressive, unconventional play and made the game more exciting than it had ever been. On why humans still bother, Allebest is blunt: "humans want to do human stuff." On acquiring expertise, his data across millions of players yields an answer he calls fortunate and unfortunate, which is that there are no shortcuts to repetition, only shortcuts in the tools. His framing for that: take any given day of your life and multiply it by a thousand, because that is what your life looks like. The company is now applying the same ratings playbook to poker at gambit.com, on the theory that people will care about a rating as much as money because it measures skill rather than who can rebuy the most chips.

• #products
• #evals

Batching agent questions like speculative decoding, and it works for design work

Human input is the expensive part of an agent loop, so instead of round-by-round questions, he built an /align-me modification that asks in batches. The analogy is speculative decoding: you gain speed by looking ahead 2 to 10 steps rather than waiting for each confirmation. He reports it works exceptionally well for design exploration specifically.

• #agents

Two categories of startup left: AI-native rocketship or left for dead

The middle has vanished. Being a rocketship means permanent fundraising at increasingly scary valuations just to hoover up capital and talent, plus a death match with other rocketships for customers that eats your gross margins. Everything else is left for dead regardless of how good the company actually is. This has been true for a while, but Turck reads the last stretch as escalating it another notch.

• #funding

One master agent or many specialized bots? Kothari expects the orchestrator wins

Grok Bot's design treats agents like people, compartmentalizing tools, context, and outcomes per bot so you go to a specific one for a specific goal. The alternative is the Jarvis model, a single master agent that just gets the job done and spawns whatever sub-bots it needs, turning the interface from 1:many into 1:1. His guess is the field starts narrow because of real engineering constraints around context windows, tool use, and cost, then converges on the master orchestrator that reads and understands the bots you create.

• #agents
• #products

Google Docs, Sheets, and Slides now work inside ChatGPT

You can open a doc in ChatGPT and chat or talk your way through changes without leaving the flow, which he says has changed how he writes, brainstorms, and proofreads. Separately, a Computer History plugin that roasts your day of computer usage turned out to be genuinely diagnostic: 48% of one day's recorded activity was Slack, Delete was the most-used shortcut at 1,191 presses, and the busiest hour was 9pm.

• #products

Every job title in demand right now still has "engineer" in it

The prediction was that AI coding would devalue engineers. The observed labor market says otherwise: forward-deployed engineer, design engineer, product engineer, growth engineer. The word survived and multiplied into new roles rather than disappearing.

• #products

AI's most useful role in a family health crisis was fighting the bureaucracy

Going into a family health situation, the expectation was using AI to research and understand the illness. What actually helped most was navigating healthcare system bureaucracy. He is also asking an open question worth watching: whether product specs now need two separate sections, one for humans and one for agents, or whether one document can serve both.

• #products
• #agents