The Agent Engineering Playbook

The Agent Engineering Playbook

Five of this week’s saves say the same thing from five angles: the agent era has stopped being about prompts and started being about engineering. Karpathy released a free two-hour lecture on agents — loops, the harness, self-improving systems — and the rest of your pile fills in the discipline around it: how to build the environment that makes agents work, how to stop them bloating your codebase, and where the economic value actually lands. The prompt is the demo. The loop is the product.

Karpathy’s lecture arrives at the moment his own experiment went from demo to discipline. His open-source autoresearch project (March 2026) put an agent in charge of improving single-GPU LLM training: the agent edits train.py, runs a fixed five-minute experiment, scores validation bits-per-byte, keeps what improves, and iterates. Seven hundred experiments over two nights, twenty optimisations discovered — greedy hill-climbing over code edits with a fixed compute budget so every run is comparable. The loop, in its cleanest form: Literature → Hypothesis → Code → Execute → Analyse → Iterate. Researchers have since shown the gains come from loop structure, not model capability — the same model at the meta level, five times the improvement, because the outer loop rewrites search code instead of tuning parameters.

Your other saves are the operating manual for that loop. @0xwhrrari frames it as harness engineering: the job is to build the environment — sandbox, tools, a fixed clock, an honest fitness function — that makes the agent’s iterations comparable and its wins cumulative. @championswimmer supplies the sharpest negative lesson: prompt skills and rules files cannot stop code bloat, because the agent that wrote the mess can no longer see it. The fix is a fresh-context review agent — a second pass with clean eyes, the way a senior dev reviews a junior’s PR. And Karpathy’s own coding workflow, broken down this month, distils to three guardrails your saves keep circling: think before coding (state assumptions before touching the codebase), simplicity first (if fifty lines solve it, never let the agent write two hundred), and surgical changes (no drive-by refactors). The pattern across all of them: the agent behaves like a hasty, over-confident junior developer, so you manage it like one — with process, not hope.

Then FarzaTV answers the “so what” twice. First, the killer use-case for personal agents: not answering questions, but helping individuals create economic value — agents as leverage for one person’s output, which is the demand story for everything above. Second, the product shape that follows: collaborators, mini-AIs that help you build, find users, and earn your first dollar. It’s the same arc as the Harsha Kumar essay from Edition 3 — the assistant that acts — now with an engineering manual attached.

• The loop is the product. Comparable experiments (fixed budget, honest metric) plus a learnings store beat a smarter model every time.

• Manage the agent like a junior dev. State assumptions, demand simplicity, forbid drive-by refactors — in process, not in prompts.

• Fresh context beats more context. A second agent with clean eyes catches bloat the author-agent is blind to.

• Economic value is the killer use-case. Personal agents win by making one person produce more, not by chatting better.

Originals: Karpathy’s 2-hour lecture · FarzaTV: collaborators · FarzaTV: agents & economic value · 0xwhrrari: harness engineering · championswimmer: fresh-context review

Capability is sprinting. Adoption is walking.

@akothari shared a Pat Grady slide making the cleanest point in this edition: the gap between what AI can do and what organisations have adopted is enormous — and gaps that big are where fortunes are made. @waiting4_asi puts a clock on it: a three-to-four-month AI arbitrage window, in which whoever exploits the niches first wins fast. And @PromptLLM shows the frontier moving again — an image post on how Opus 5.5 would beat everyone else — which only widens the gap further.

The mechanism is simple: capability compounds on research timelines while adoption compounds on organisational timelines — budgets, habits, procurement. Every model release re-opens the spread. The arbitrage window is the period between “the model can do it” and “everyone’s procurement approved it”. In that window, a small team with no approval process can ship what an enterprise is still forming a committee about. Three to four months sounds short; it’s roughly how long the gap has been open, on repeat, for two years.

• Watch the gap, not the leaderboard. The money is in the spread between capability and adoption, which every release refreshes.

• Speed is the moat inside the window. Small teams win arbitrage plays because they skip the committee.

• Niches first. The window rewards whoever picks an unguarded niche and exploits it fast, not whoever builds the biggest model.

Originals: Pat Grady’s slide · the 3–4 month arbitrage window · how Opus 5.5 would beat everyone

The $5 Trillion AI Roll-Up

Greg Isenberg’s thesis is the capability gap, weaponised: buy boring businesses, then triple their EBITDA with agents. The private-equity roll-up playbook — acquire at low multiples, centralise, cut costs — has always worked; AI agents just made the “cut costs” step an order of magnitude more violent, and available to anyone with a laptop instead of a fund.

Why this is a $5T idea and not a $5B one: the target set is every service business that never got the software treatment — the plumbers’ dispatchers, the clinics’ schedulers, the logistics brokers. Software ate the front office; agents are eating the back office. And the buyer doesn’t need to be a fund anymore: the harness from this edition’s lead story is the operating partner.

Original: $5T AI roll-ups: buy boring businesses, 3× EBITDA with agents

Notes on Sequoia’s Playbook

Notes circulating on how Sequoia thinks — the firm that backed Shaadi.com back in 2006 and has ridden every one of these cycles. Read it as the institutional version of the arbitrage thesis: the best firms don’t predict the gap, they position inside it early and let the compounding do the work.

Original: Notes on Sequoia’s…

The AI Update, Boardroom Edition

A Loom recording of an AI update delivered to the Boston College investment committee — the adoption side of the capability gap, in the wild. Endowments move slowly; the fact that “AI update” is now standing agenda at investment committees tells you the gap is closing from the demand side too.

Original: Loom: AI update for the Boston College investment committee

How to Level Up an Org, in Three Phases

Slava Akhmechet’s framework for levelling up organisations runs in three phases — learning, changes, elevation. Phase one is pure observation: learn how the org actually works, who really decides, where the energy is. Phase two is making the changes the learning revealed — structure, process, the hard calls. Phase three is the payoff most leaders skip: elevate the hidden stars, the people doing the real work two levels below where the credit lands. The framework rhymes with the agent loop from the lead: observe (literature), act (code), keep what works (analyse) — management as hill-climbing.

@vijayshekhar adds the cultural footnote: CEOs speaking their mother tongue — the signal that leadership has stopped performing for someone else’s template. It reads as the human version of phase three: authenticity as elevation, the org finally speaking in its own voice.

Originals: how to level up orgs · CEOs speaking mother tongue

The Shaadi.com Origin Nobody Remembers

Before it was Anupam Mittal’s sagaai.com, shaadi.com was two guys, a brand-new internet, and a domain bought from Network Solutions — the only place you could buy one. Manish Jain’s thirtieth-anniversary post is the best founder story in this edition: no grand plan, no market research, just a name that felt obvious and a static site built in Microsoft FrontPage.

• Obvious names are underpriced. “A name that felt obvious” beat every branding exercise — twice, by two different founders.

• Timing is a feature. The Hotmail accident — 35% of members arriving on an 89-day-old email service — is the whole 1996 internet in one stat.

• Bootstrap until the brand exists. Mittal waited nine years and a real business before taking Sequoia’s money in 2006.

Original: thread on Shaadi.com’s origin · Manish Jain’s anniversary post: celestri.org

Paul Graham’s Best Essay, in 200 Seconds

Deedy Das keeps proving the same point: a single essay plus Claude is now a film studio. This week’s exhibits — Paul Graham’s “How to Do Great Work” compressed into a 200-second AI video, and the entire history of Google told in two minutes, both generated with Claude.

The deeper point: the binding constraint on educational media was never cameras or editing — it was the compression of a long text into a tight visual script, and narration, and pacing. All of that is now one model’s afternoon. Deedy’s fully-automated YouTuber pipeline from Edition 3 was the factory; these two videos are the product line running. Expect the “200-second version” of every canonical essay within a year — and expect the winners to be the ones with taste in what to compress, not access to the tools.

Originals: PG’s “How to Do Great Work” as a 200-second video · Google’s history in 2 minutes, via Claude

Tata Tests Where It Matters

Tata’s ad — cars tested in real Indian conditions — went viral, and the reason is structural: Indian roads are the world’s hardest durability lab. @nitprashant makes the companion point: India’s worst traffic doubles as the ultimate car durability test. Marketing that shows the actual test, on the actual roads, beats a studio shoot because the audience drives those roads.

Originals: Tata’s viral real-conditions ad · traffic as a durability test

The Mother-Tongue CEO

Filed under organisations above, but worth its own beat: the Indian founder who stops presenting in borrowed language. As India’s startup voice matures, the mother tongue isn’t folksy — it’s precision. You argue best in the language you think in.

Original: CEOs speaking mother tongue