OpenAI Says Half Its 4-to-8-Hour Agent Tasks Needed a Human to Step In. That Is the Number That Matters.

OpenAI published a post on September 6 titled "Research acceleration: The view inside OpenAI" announcing that it has hit the first of the two milestones Sam Altman set out in October 2025: an automated research intern, defined as "a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days." The second milestone, an automated AI researcher, is still pencilled in for March 2028.

The numbers underneath the headline are more useful than the headline.

What OpenAI actually measured

The post is unusual for OpenAI in that it reports internal operational metrics rather than benchmark scores. According to the figures quoted in AI Weekly's summary and OfficeChai's writeup:

• The research org now runs about 3.1 agent-workdays for every human workday, measured against an eight-hour day. Before June 2026 total agent runtime was still below total human labor hours, so this crossover is roughly three months old.
• The median researcher, by agent usage, was spending more than $600 a day on inference at API prices by mid-August. The heaviest users were north of $7,000 a day.
• Over the last six months, more than half of successful tasks estimated at four to eight hours of human work needed at least one human intervention along the way.

All of these are OpenAI's own internal measurements. Nobody outside has verified them. The third bullet is the one worth sitting with, because it describes the operating reality of the "intern" far better than the milestone language does.

Half the long tasks needed someone to step in

A task that would take a person a full working day, run by an agent, and in most successful cases a human had to intervene at least once. That is not a criticism of the model. It is a description of what a multi-hour autonomous job looks like in practice today, in the most agent-saturated research organization on the planet.

It also tells you what the bottleneck is. If the agent runs for six hours and a human has to notice that it went sideways at hour three, the question that decides your throughput is not how smart the agent is. It is how quickly the human finds out. The post itself says humans still choose priorities, judge results, intervene on difficult tasks, and decide whether to scale. Every one of those verbs needs a place to look.

OpenAI's post does not describe its supervision tooling. Outside a lab with a dedicated infra team, most people running an agent for six hours are watching a terminal on one machine, and the moment they walk away from that machine the intervention loop breaks.

The supervision channel is a terminal, so make it reachable

This is the problem shell.online was built for, so it is worth being precise about what it does.

The CLI wraps a command in a local PTY, keeps the process on your machine, and prints a URL plus an eight-character password that opens the same live terminal from a phone or another laptop. It is the real session rendering as it runs, not a log tail or a screenshot. For a long-running research job the relevant flags are the ones in the CLI reference in the repo:

• shell --read-only python train.py creates a view-only link. Read-only is enforced by the relay Worker, not the browser: ordinary input and EOF are rejected server-side, so DevTools or a handcrafted WebSocket frame cannot turn it interactive. Access mode cannot be weakened after creation.
• shell --json -- <command> emits one JSON event on stderr with the share URL and, for encrypted sessions, the E2EE password, while the wrapped process keeps stdout. That is the path for an agent to create its own share unattended and hand the link back to whoever is supervising it. There is an installable agent skill for exactly that.
• shell --persistent <state-file> <command> stores an owner-only host identity, password, and E2EE material so the same URL survives a process or machine restart. If the intern crashes at hour four and gets relaunched, the link the human already has still works.

Terminal frames are end-to-end encrypted by default. The CLI encrypts before Cloudflare relays them and the browser decrypts locally, so the relay sees connection and lifecycle metadata but not terminal input or output. The security model spells out what the relay can and cannot observe. If the host drops off the network, the same ordinary share link stays recoverable for twelve hours, and the share closes automatically when the wrapped process exits.

None of that makes the agent better. It makes the "at least one intervention" cheaper, and at 3.1 agent-workdays per human day that is the number that moves throughput.

The next milestone changes the shape of the problem

The March 2028 target is an automated AI researcher, not an intern. The difference OpenAI draws is direction: the intern works under human direction on well-defined tasks. A researcher picks the task. That implies agents delegating to other agents, and OpenAI's own caveat is blunt: "we do not yet know how to safely get all the way to aligned, full RSI" and "we cannot assume that progress in alignment and safety will keep pace." The post commits to slowing down or halting work whenever a system is judged too risky to monitor adequately.

Monitoring is the operative word again, and once agents are talking to each other the monitoring surface is no longer a single terminal. It is the network between them. This is the layer Pilot Protocol sits at. Every agent gets a 48-bit virtual address and an Ed25519 identity bound to its tunnel sessions, nodes are private by default, and a peer can only be reached after a signed mutual trust handshake. The layers.yaml in the repo declares the stack one Go package per layer, with imports only allowed downward: key exchange at L5, the AEAD envelope and replay window at L6, the reliable stream at L7. Doing it as a network layer rather than a platform API means reachability is decided by signed trust state, not by whatever the model decides to call.

The numbers in this post are more candid than most labs publish. Read plainly, they say that autonomy at the multi-hour scale is currently a human-in-the-loop system with a very long loop. The tooling that shortens the loop is unglamorous, and it is where the throughput is.