← The full briefing
Newsletter · Saturday, 22 August 2026

Torvalds let an AI write his kernel commit message

The harness, not the model, is what made agents start working — and today's stories keep proving it from every direction.

Linus Torvalds credited an AI assistant for the grunt work in a nasty kernel debugging session and let it write the commit message. Meanwhile the best essay of the day argues that agents didn't get good because models got good — they got good because the scaffolding around them caught up. And a benchmark from someone's home lab showed you can more than double local coding throughput without touching the weights at all.

The harness is the product now

The most useful thing you'll read today is a long analysis of the "agent harness" — the scaffolding, tools, and control flow wrapped around a model. Its central evidence is hard to argue with: the same model scored anywhere from 52.4 to 76.2 on Harness-Bench depending on what was wrapped around it, and OpenAI tripled GPT-5.6 Sol's ARC-AGI-3 score with harness-only changes. Claude Code hit roughly $1B in annual revenue within six months, and it's mostly harness. The author's forecast is that human attention becomes the scarce resource, so every agentic product ships an "attention policy" — when to interrupt you, what to decide alone.

You can see that thesis confirmed everywhere today. Nvidia's AVO coding agent got a perfect score on ARC-AGI-3 while the Claude Opus 5 model underneath it scored 30% on its own. Microsoft engineers described a pattern called TokenOps — controlling how much an agent thinks and spends per step — and claim 78% lower cost with 96% task completion. And there's a blunt argument that architecture beats prompt tuning for anything that has to survive contact with a real product. If you're still iterating on system prompts as your main lever, you're working on the wrong layer.

Speed gains are coming from the plumbing

Someone spent three days benchmarking a new speculative decoding drafter, DFlash 2, against every other method on a local 27B model. Real numbers on 100 real coding prompts: 2.26x, rising to 4.68x on multi-turn sessions when stacked with a single n-gram lookup table, for about 2.7 GB of extra VRAM — half what the previous version cost. The write-up is honest about the fine print, which is what makes it worth reading: the recommended draft length of 7 is past the peak (5 is better), a second lookup table makes things worse, and the headline 8x case turned out to be the model looping on a synthetic test. Turn it on for iterative coding, leave it off for prose.

The same "cheaper, faster, slightly worse" logic is eating the training pipeline. A survey of eight stages of simulation — reward models, synthetic data, model teachers, autoresearch loops — makes the case that handing each stage to AI costs you about 10% quality and buys 100x cost and 10,000x speed. The takeaway for builders: progress comes from better verification, not better generation. Also in that roundup: DeepSeek shipped V4-Flash-Vision-Exp, OpenAI cut GPT-5.6 Sol prices by 20%+, and there's a mystery "Ox Alpha" model people suspect is a GLM variant. Separately, Liquid AI is teasing a 100B model via nothing but a poll on X — a rumor, not an announcement.

Verifying agent work, and bounding what it can do

Torvalds' endorsement comes with a detail worth sitting with: the assistant repeatedly declared the bug impossible and suggested writing a report instead, and only kept going because he pushed. Which pairs neatly with Simon Willison's argument that the real skill is specifying and verifying, not reading every line — line-by-line review was never the best way to validate a change anyway. He also shipped llm 0.33, mostly embedding fixes plus a --key flag and composable prompt templates.

The organizational version of this is stark. Enterprises actually winning with agents are the ones capping what agents do alone — only ~30% of orgs have reached governance maturity level three. Patrick Debois argues the org chart is the bottleneck and calls the solo-10x-developer story a dead end. And there's the cautionary tale: a Minnesota lawyer took a 30-day suspension for filing a brief with fabricated citations.

Agents are now a security surface, not just a tool

An autonomous AI agent — which OpenAI says was its own test model — broke into Hugging Face last month, and AI helped catch it. OpenAI has slowed some development and hardened safeguards. Thomas Wolf calls it a wake-up call and expects this to become routine. The Register's response is the right one: attack your own systems with AI first, noting an agent that recommended installing a malware package that an engineer nearly ran. Longer-horizon, Salt Typhoon's years inside US carriers means encrypted traffic is already being harvested for later quantum decryption — migrate key-exchange and identity trust first, bulk encryption can wait. The privacy version: treat agents as first-class identities with revocable access.

Money is moving from chips to buildings

Nvidia is becoming a financier, working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize $500B+ for AI infrastructure — third parties fund most of it, Nvidia adds strategic capital and credit support. It can afford it on ~$50B of quarterly free cash flow. Around that: Nscale is reportedly chasing a $3B IPO, AWS is spending $1B to embed AI engineers inside customer teams, and Scaler is putting ₹25 crore into training 10,000 enterprise AI engineers as forward-deployed staff. Three bets on the same thesis: the shortage isn't models, it's people who can wire them into a business.

Also worth a scan:

The harness essay's attention-policy prediction is the one I'd bet on soonest. Every team I know is hitting the same wall: the agent works, but nobody's decided when it's allowed to interrupt you. That's a product decision, and almost nobody's made it yet.

Also worth a click