Assistance vs Automation: Why AI Feels Jagged

Two things are true at the same time, and the field keeps refusing to hold both. Models post record scores on benchmark after benchmark. And outside a few demos, most work looks much as it did three years ago: the same approvals, the same spreadsheets, the same person copying one number from one system into another. If AI is as transformative as the benchmarks say, where is all the automation? The sharpest answer going comes down to a distinction most teams never make: assistance vs automation.
Diogo Almeida — a co-author of the InstructGPT paper and now a co-founder of TypeSafe AI — put that question at the centre of a talk at AI Engineer published on 15 November 2026, and refused the usual exits. Not "it is early". Not "the next model fixes it". He wanted an Occam's razor answer to why AI is superhuman at some tasks and useless at the ones next door. His answer is that distinction, and it explains more than it has any right to.
TL;DR — Almeida's answer: AI looks jagged because nearly every frontier model was optimized for assistance, where a person reads the output, and almost none for automation, where nobody is watching. On his account the human was in the training objective by construction: the HF in RLHF is a reader whose approval the model learns to maximise. That makes 95% accuracy comfortable for a draft and fatal for a chain of ten unattended steps. Running work unattended means pulling the two things the reviewer used to supply — the missing context, and the judgment about when to escalate — out of the chat window and into systems that can fetch and record them. The context half is what a markdown workspace like MDflow is for. Start free.
Assistance vs automation: what is the difference?
Assistance is work a person reads before it counts. Automation is work that counts without anyone reading it. The model can be identical in both cases. What changes is who absorbs the mistakes, and where the missing context comes from.
| Assistance | Automation | |
|---|---|---|
| Who reads the output | a person, every time | nobody, usually |
| What a good output optimizes for | being useful to that reader | being right, in a shape code can act on |
| Where missing context comes from | the person, mid-conversation | whatever the system can fetch itself |
| What one bad output costs | a correction | a wrong row, downstream, silently |
| What variance does | annoys you | ends the run |
Almeida's own example is deep research. He says he could not write a better research report than the tools can — but if the answer simply had to be correct, rather than well written, he would back himself over them every time. That gap is invisible in assistance, because the reader closes it for free. It is the whole problem in automation.
And notice what the reviewer silently provides besides error correction: everything the task needed that nobody wrote down. The constraint from last quarter. The reason that obvious approach was rejected. The fact that this customer is on a legacy plan. In assistance the model does not need to retrieve any of it, because a human is sitting there to be asked. Remove the human and that context does not relocate itself — it just goes missing, which is the same failure mode we described in Implicit Knowledge Is Invisible to AI Agents.
Why frontier models are optimized for assistance
Because a human reader is in the training objective, by construction. Reinforcement learning from human feedback collects human rankings of model outputs, trains a reward model on them, then optimizes the model against that reward. The quantity being maximised is a reader's approval. Almeida's claim — and he is unusually well placed to make it, as a co-author of the paper that made the technique famous — is that this is precisely why today's models still need a human in the loop. The HF in RLHF is not an implementation detail. It is the spec.
That is his argument rather than an established finding, and he is selling an alternative, so weigh it accordingly. But the same paper supplies the evidence for the other half of it. In the InstructGPT human evaluations, outputs from the 1.3B-parameter model were preferred to outputs from 175B GPT-3, with roughly 100x fewer parameters. A change in what the model was trained to do beat a hundredfold difference in scale.
Almeida stacks that into a ladder. Rich Sutton's bitter lesson says scale beats clever algorithms. The bitterer lesson, he argues, is that data beats compute. The bitterest is that doing the right task beats both — and no one has yet done the work of optimizing seriously for the task of deciding unattended. You get what you optimize for, and for four years the field has optimized for being pleasant to talk to.
Why reliability, not intelligence, is the blocker
Almeida's line for this is the best in the talk:
"You get Einstein 95% of the time, but Mr. Bean 5% of the time."
The point is about variance, not the average. Assistance tolerates 95/5 beautifully, because the reader catches Mr. Bean the moment he appears and the cost is one regenerate. Automation multiplies it: a 95% step chained ten times, with no one looking in between, completes correctly about 60% of the time. And if you cannot tell at the moment of output which of the two answered, there is no threshold to set and nothing to escalate — which is the real blocker, not raw capability.
It also explains a side effect everyone complains about. Almeida's account of slop is an economic one: because AI output is unreliable, you cannot use it for decisions that cost you anything — but you can still push the cost onto other people. Unreviewed output is cheap to produce and expensive to receive. Reliability is what makes the generator bear its own costs.
What a model built for automation looks like
TypeSafe's bet is the obvious consequence of the diagnosis: if models are bad at automation because nobody optimized for it, optimize for it. Their first model, Jev, was covered in detail by InfoQ on 1 October 2026, and the shape matters more than the benchmark:
- It does not generate text. A caller sends a state plus a set of typed questions; the model answers them in a single parallel pass, returning typed answers — Choice and Score among them — each with a probability distribution and a confidence value.
- That confidence is the product. Calling code acts above a threshold and escalates below it, which is the control flow assistance never needed.
- TypeSafe quotes end-to-end latency of 70 to 500 ms and trains with a method it calls Reinforcement Learning for Calibrated Decisions. Vercel added it to AI Gateway on day two and reported it reaching nearly 13% of paid teams within 24 hours; LangChain shipped middleware that screens tool calls before they run.
The caveats in the same report are as instructive as the claims. A developer on Hacker News pointed out that the model cannot emit an invalid type but can still emit a completely wrong valid value. Armin Ronacher, CTO of Earendil, told TechCrunch the design hands part of the hallucination problem back to the caller, who now has to decide whether a 50% probability is worth acting on. And an analysis of 12,759 launch posts put user-reported speedups at a median of 7x against a 193.6x headline.
The most revealing artefact is the vendor's own "jaggedness" page, which documents that the model is unreliable at counting, arithmetic and date comparison, and advises keeping that logic in code. Read as an admission, it is damning. Read as a design, it is the entire thesis: the model decides, code computes. Almeida's framing in the talk is that code is the language of automation, so the useful thing an AI can do is make the judgment call and hand control straight back to software — the same instinct behind state machines for AI agents and splitting the task from the model.
Which applications benefit most
The test for each is the same: would anyone read this output if the system did not flag it?
- Routing and classification — which queue, which model, which team. High volume, no reader, and a confidence number is directly actionable.
- Guardrails on tool calls — screening an action before it runs, where latency is a hard budget and a paragraph of reasoning is useless.
- Triage — support inboxes, alerts, inbound leads. The decision is "does a human need to see this", which is exactly a threshold.
- Extraction gating — not pulling the value out, but deciding whether this document is the right one and whether the match is good enough to act on. See Document Extraction for AI Agents.
- Scheduled background work — anything that runs at 03:00 and has to be either right or visibly escalated by morning.
- Monitoring and drift checks — judging whether this week's distribution still looks like last week's, as in Agent Drift Detection.
What all six share is that nobody is reading the output, so the cost of a wrong answer lands on someone downstream who never asked for it.
How MDflow fits
Start with the boundary. MDflow is not a decision model. It does not score anything, returns no confidence value, and has no opinion about when your agent should escalate. That judgment layer belongs in your code, and the quickest way to misread this post would be to think a knowledge base makes automation reliable.
What it does supply is the other thing the reviewer used to carry: the context, fetchable without anybody pasting it.
What lines up today
- Retrieval by purpose, not by filename. A folder description is the primary ranking signal, and
mdflow_get_contextranks those descriptions above folder names and document titles. A folder described as "refund rules and the exceptions we have agreed" is what an unattended run finds when it asks about refunds — the mechanism we cover in Folder Descriptions as Agent Context. - Read and write, over MCP and HTTP. The MCP server and HTTP API both write, so a scheduled run can read its brief and file what it decided into the same workspace, rather than leaving the record in a chat nobody will reopen. See Let AI Agents Write to Your Knowledge Base.
- A record of what ran while you were not looking. In a personal workspace, the Document Log at
/logrecords created, edited, shared and deleted events with the actor on every row —youfor browser actions,automated · <token name>for anything arriving over the API or MCP — and an edited row opens a diff of exactly what changed, retained for 30 days. Team accounts keep a separate audit feed of account, membership, workspace, folder and document lifecycle changes, retained for 365 days. - Version history that covers machine writes. Every saved body change keeps the previous version across every write path, including the API and the MCP server, with a line-by-line diff and a non-destructive restore. An unattended agent that writes something wrong leaves the previous version behind.
Connect Claude to MDflow: open the MDflow connector in Claude's Connectors Directory, click Connect to Claude, and sign in — no URL or token to paste. One connection covers claude.ai, the Claude desktop and mobile apps, and Claude Code signed in with a Claude subscription. Team workspaces use an
mdt_Team access token instead, with the bearer-header clients listed in the MCP docs.
Where the report card is worse. Three gaps that matter specifically for unattended work:
- A personal access token is not scoped. An
mdf_token is account-wide and read and write, with no read-only option and no per-workspace restriction. Only Team tokens (mdt_) are bound to named workspaces and can be capped read-only. If an agent is going to act while you sleep, size that blast radius deliberately — the argument in Agent Authorization. - Version history is not reachable by an agent. It is visible in the editor — active Team members can compare and restore Team document versions too — but it is never exposed over the API or MCP, so an agent cannot ask what changed since its last run. Retention is bounded either way, at 30 days or the last 300 versions per document. On a personal document it is narrower still: history is private to the owner, and it is a Pro feature.
- Writing the context down is still a human step. Nothing distils last quarter's decision into the folder for you, and that is the main reason unattended runs still arrive underbriefed — in this tool as in every other.
Where we are headed — direction, not a dated commitment. The folder description stays the primitive we invest in, because stating what a set of documents is for is what makes retrieval work without an index, and an unattended agent cannot ask a follow-up question. The other thread is shrinking the distance between where a decision gets made and where it gets written down.
The bottom line
On Almeida's account the jagged frontier is not a mystery about intelligence. It is a mismatch between what models were trained to do and what we are asking them to do. Assistance was optimized for; automation mostly was not, and in automation the reviewer who used to absorb the errors and supply the missing context is simply absent.
That gives you two jobs, neither of which is waiting for a better model. Put the judgment in code — a threshold, an escalation path, arithmetic that a model never touches. And put the context somewhere an agent can reach on its own credentials, so the run at 03:00 knows what the person at the keyboard would have known.
Almeida's closing bet is that in six months or a year the question stops being can this be automated and becomes how much work is it. The parts you can do before then are unglamorous and available today.
Start free, connect an AI agent, or read the API docs.
Frequently asked questions
What is the difference between assistance and automation in AI?
Assistance is work a person reads before it counts. Automation is work that counts without anyone reading it. The model can be identical in both cases; what changes is who absorbs the mistakes. In assistance the reviewer is a free error-correction layer and also the carrier of missing context, because they can be asked a question mid-task. In automation there is no reviewer, so both the errors and the context have to be handled by the system itself.
Why are large language models optimized for assistance rather than automation?
Almeida's argument is that the human was written into the training objective. Reinforcement learning from human feedback collects human rankings of model outputs and trains a reward model on them, so the thing being maximised is a reader's approval. Diogo Almeida, a co-author of the InstructGPT paper and now a co-founder of TypeSafe AI, argues in a talk published on 15 November 2026 that this is exactly why today's models still need a human in the loop, and that pleasing a reader has little to do with getting a task right unattended.
What is a System One model?
System One Model is TypeSafe AI's term for a model that returns typed decisions instead of generated text. According to InfoQ's report of the Jev release on 1 October 2026, a caller sends a state plus a set of typed questions, and the model answers them in a single parallel pass with a probability distribution and a confidence value, so the calling code can act above a threshold and escalate below it. TypeSafe quotes end-to-end latency of 70 to 500 milliseconds.
Why does 95% accuracy break automation?
Because automation chains steps and assistance does not. A step that is right 95% of the time, repeated ten times with no reviewer in between, finishes the chain correctly about 60% of the time. In assistance that same 95% is comfortable, because a person catches the bad one as it appears. The number that matters for unattended work is not average accuracy but whether you can tell, at the time, which answers to trust.
What does an AI agent need in order to run unattended?
Three things the chat window usually supplies by hand: the context it needs without anyone pasting it, a way to record what it did, and a defined action for the cases where it is not confident. The first is a retrieval problem — the brief, the constraints and the prior decisions have to live somewhere the agent can fetch on its own credentials. The third is why calibrated confidence matters: unattended work needs a threshold and an escalation path, not a better paragraph.
Further reading
- Diogo Almeida, "Where is all the automation?" — the AI Engineer talk this post is built on, published 15 November 2026.
- Training language models to follow instructions with human feedback — the InstructGPT paper (Ouyang et al., 2022), source of the 1.3B-versus-175B result.
- Rich Sutton, "The Bitter Lesson" — the 2019 essay the ladder starts from.
- InfoQ: TypeSafe AI releases Jev — the release report, including the adoption numbers and the critiques.
- Context Engineering for AI Agents — how to decide what fills the window when no one is there to add to it.
- Quality Gates for AI Agents — what to put in code when you cannot rely on a reviewer.
- MDflow MCP docs and the FAQ.