A looped transformer reuses its layers to do more computation. Experiments on NVIDIA’s Dory suggest that the gains can disproportionately favor reasoning without a written chain of thought.
Read the essayThe Orchestrator · Field notes
The book ships weekly. The notes ship faster.
Essays on agent platforms, delegated work, and the governance layer nobody has shipped yet — notes from the writing of The Orchestrator.
Newsletter
New essays and chapter updates, a few times a month. No spam — unsubscribe any time.
More from the blog
Anthropic just made Advisor and Orchestrator production features. You do not need Claude Code - or Claude - to get the same effect: a router that knows what things cost, and an on-call expert for when the cheap model is stuck.
Read the essay
30% faster, 20% cheaper, under GPT-6 Astra - the model numbers are good. The game changers are Advisor and Orchestrator: dynamic model selection that cuts the waste out of intelligent multi-model workflows. I once spent $1,800 on a single premium-model session to learn what the alternative costs.
Read the essay
Empero's Manta reads a hundred million tokens in one prompt and found 18 of 20 hidden codes in 88 million tokens of SEC filings, with the data on GitHub. The price is blank — and on today's arithmetic, it would be a big one.
Read the essayOpenAI asked a Microsoft researcher where one of his ideas came from. The idea: AI models shred their own thinking between every word — and a fix exists.
Read the essay
Moonshot shipped 1.56 TB of open weights with a license that reads like MIT until you hit Section 3 — and the questions it raises about attribution and enforcement are the interesting part.
Read the essayI was always skeptical of skills. Now people tell me models are powerful enough that you should get out of their way. Both sides are half right, and one side is doing arithmetic wr
Read the essayJEPA predicts representations instead of generating pixels or tokens. The checkpoints are public, one descendant is already in production, and it is the architecture worth tracking
Read the essay
Within an hour of the O'Reilly Radar republication, collaborators started writing. Now the drafty preprints have to become commitments — and the distinction the conference kept collapsing, semantic versus epistemic governance, is worth restating.
Read the essay
Within an hour of the O'Reilly Radar republication, collaborators started writing. Now the drafty preprints have to become commitments — and the distinction the conference kept collapsing, semantic versus epistemic governance, is worth restating.
Read the essay
You can hear whether someone is rushed, angry, or quietly dissatisfied. A keyboard flattens all of it. Prosody is the second channel inside voice feedback, and systems that react to it are the undiscovered country.
Read the essay
The nightly pricing check came back “no change” almost every night this month. The stable sticker prices are hiding where the action went: new tiers underneath the flagships, and a consumption mix tipping decisively toward cheap models.
Read the essay
The most useful part of Anthropic's S-1 is not the financials. It is the platform section: what a harness is, which primitives exist, and exactly who owns which decisions — the vendor runs the loop, the orchestrator supplies the intent, the tools, and the guardrails. Itemized under securities liability.
Read the essay
Five times more gets done talking to a series of agents than typing the same instructions. The tools are here — Claude Code voice mode, Muesli, the dictation layer. The return-to-office mandate was designed for the keyboard.
Read the essay
Steve Yegge retired Gas Town in August — his own words: it effectively burned down. The launch reached tens of thousands; the retirement note barely traveled. What people keep using matters more than what launches well.
Read the essay
An orchestrator spends most of its decision budget not writing — it gates. Jeff, a 0.8B decision model, beat a 27B on those calls by eight points at 35x the speed, for about two gigabytes of memory. The big model does the thinking; the small one makes the calls at the door.
Read the essay
Installing the tools everyone else is installing is the airstrip, not the airplane. These models aren't just for writing code — they're for thinking about whether you should write it at all. Run Feynman's test on your week: did any design decision change, or only the typing?
Read the essay
A dashboard is a picture of a dataset. The consumer of your data is no longer a person at a screen — it's an agent that needs the data itself: where it is, what it means, and enough context to act. Build the dataset first.
Read the essay
A chat window solves the problem you typed. Orchestration builds the loop that answers the question you were about to type — and the one after that — before anyone asks. The specific request is no longer the unit of work. The loop is.
Read the essayThe first results from the governance benchmark say that stronger models handle semantic and epistemic puzzles better than weaker ones. Nobody will fight that conclusion. The work
Read the essay
Why a benchmark needs to check both the consequences of a change and the evidence used to justify it. The idea behind Paper 1.
Read the essay
A plain-language walkthrough of rule changes, conflicting evidence, and the difference between reflection and checker-assisted correction.
Read the essay
The corrected ten-run results from six hosted AI configurations, what varies between judgments, and what the Basic Level 1 tests establish.
Read the essay
Frontier models follow long, contradictory rule systems almost perfectly. Hand them a pile of semi-trustworthy sources instead and something strange happens: high reasoning effort scores better than extra-high. More thinking, worse answers — the complexity limit.
Read the essay
The frontier models follow long, contradictory rule systems almost perfectly. The budget tiers don't. The gap between them is a compliance cliff, it concentrates where stakes are highest, and the numbers that measure it aren't ready yet.
Read the essay
I run a benchmark of abstract logic problems across a dozen models. One proprietary content filter called it a cyber attack. Here is the practical argument for open weights that no benchmark captures.
Read the essay
Noam Brown says alignment is the number one priority. Susan Rice wants a freeze. Both skip the step in the middle: a deterministic layer that decides what an agent is allowed to do — and allowed to conclude — no matter what it wants.
Read the essay
I spent my formative years trying to read philosophy, and words like ontology and epistemic have carried a built-in fear ever since. What epistemic actually means — and why it's the word at the center of governing what AI agents believe.
Read the essayA new preprint examines how changing a category can broaden an agent’s decision authority without changing its permissions—and how that change can be reviewed before use.
Read the essayA new architectural preprint connects changes in an agentic system’s domain model to capability development, authority review, and the decisions it may automate.
Read the essay
Every agent platform governs what an agent can do — tools, files, actions. Ontology. What almost none of them govern is what an agent can believe, conclude, or think of. Epistemology is the other permission system, and nobody has shipped it.
Read the essay
NVIDIA shipped a hardware watchdog for agents. Berkeley shipped privilege control. OWASP shipped a Top 10. Every one of them governs what an agent can do — none of them governs what it can believe, and that is where the next attack already lives.
Read the essay
Last week the rumor was that Google cracked recursive self-improvement and that an OpenAI model called Bel produced 722 math proofs. None of it confirmed. The confirmed things are quieter: the labs say self-improvement is coming faster than they thought, and nobody has written down the terms.
Read the essay
When you draft an agent, you are setting expectations. Baseball drafts players and documents get drafted — an agent gets both at once. The oldest contracts, written in clay four thousand years before the first agent, show why the writing down is the whole relationship.
Read the essay
A book about a field that changes every week can't ship finished. The free online edition absorbs the changes — last week it was the Yegge post-mortem and the DevDay wave. What that means if you buy print.
Read the essay
Last year I still read some code. These days I try not to — and that reversal is where the Rails World keynote found me. If you still spend most of your day writing code by hand, something is going to surprise you.
Read the essay
NVIDIA fenced agents in, OpenAI made them always-on, AWS put them in your VPC — and the father of agent orchestration shelved his own orchestrator. What shipped between September 23 and October 1.
Read the essayChapter 5 surveys the tools behind orchestration, from models and memory to permissions and durable execution. It was written during an irruption, which means the map is changing w
Read the essayChapter 4 shows what orchestration looks like in real work: where it exists now, how a feature and an incident run through it, and why skills, memory, evals, traces, permissions, a
Read the essayDorsey, Zuckerberg, and Karp are all pointing at the same transition. None of them are wrong. None of them are giving you the model either.
Read the essayChapter 3 of The Orchestrator is live. The vocabulary before the toolkit: delegated intelligence, disposition, and the permission envelope that holds a system's authority.
Read the essayThe layoffs are visible. The infrastructure debt accumulating underneath every AI-coding company is not. The transition to an orchestrator-driven economy is already underway.
Read the essayOne role, many variations. The orchestrator ecosystem includes specialists, fellows, backup support, insurance, and the concept of delegated intelligence.
Read the essayChapter 2 of The Orchestrator is live. It starts with the people already building systems that connect evidence, decisions, and action, even if their job titles haven't caught up.
Read the essayThe developer role is not dying — it is transitioning. The orchestrator is the person who directs delegated intelligence to create systems and owns the outcome.
Read the essayThe orchestrator is not the person who writes every line. The orchestrator is the person who decides what can be delegated, what must be checked, and who owns the result.
Read the essayThe orchestrator is the person directing systems, agents, and outcomes — and remaining accountable for what gets delegated.
Read the essayA free, in-progress book about the emerging post-programmer IT job title. Released chapter by chapter to start the conversation before the job titles settle.
Read the essay