Using an agent is not the same as directing one

The numbers are real; that is not the useful part

The compensation figures are not invented. Robert Half's 2026 China guide puts senior and LLM-specialized comp at ¥1M+, with AI research PhDs near ¥2M. 2026 recruiting reports put the large-model algorithm track's peak around ¥52k a month and the structural AI talent gap above 40%. The exceptional new grad clearing a million RMB exists.

But these are the visible tail of a bidding war for a few hundred people, not a floor — and, more to the point for anyone reading, none of it is something you can act on this week. You cannot will yourself a ¥2M offer, and waiting for one to confirm the market is real is not a plan. A true number you cannot act on is, for practical purposes, noise. Set it aside.

The squeeze that is real: the paper treadmill is a sieve, not a ladder

The thread's accurate cluster is the research-hiring squeeze, and the hard numbers are unambiguous. CVPR 2026 put 16,092 submissions through review and accepted 4,090 at a 25.42% rate — a 42% jump in accepted-paper volume over CVPR 2025's 2,878. ICLR 2026 submissions roughly doubled, from about 11,000 to around 27,000, requiring more than 80,000 reviews. NeurIPS has gone from 9,467 submissions in 2020 to 21,575 in 2025 at a 24.52% acceptance rate.

Read those together and the shape is clear: accepted volume grows, but submission volume grows faster, so the marginal signal of one more paper depreciates while the bar to clear screening rises. The treadmill speeds up; the destination does not move closer. This is why "publish more" is not the lever it looks like — it is a sieve that keeps narrowing, not a ladder you climb. So what is the actual lever?

The frontier is on a published timeline

The reason "can you direct AI" beats "do you use AI" is not a speculative claim — it is scheduled. OpenAI's stated roadmap targets research-intern-level AI by around September 2026 and a fully automated, "legitimate" AI researcher by roughly 2028, backed by roughly $1.4T in committed infrastructure across about 30 GW of compute; current models already operate on a roughly five-hour task horizon.

The implication is specific. The capability that turns "I use AI every day" into a depreciating asset and "I can direct an agent" into leverage is on the same kind of curve as the conference numbers above — moving, dated, not fixed. An engineer whose value is "uses the tool" is competing directly with the tool's next release. An engineer whose value is "owns the loop the tool runs inside" is not. That distinction is the entire actionable content of the thread, and the rest of this defines it.

Using an agent

Using an agent is issuing a request and reading the reply. You describe a task, paste the output into your work, and fix it by hand when it is wrong. The skill ceiling is prompt phrasing. It is genuinely useful and it is where the great majority of "I use AI every day" engineers stop. Nothing in this mode survives scale: the agent does not remember the last task, cannot act on your systems, has no notion of being right or wrong, and cannot be trusted to run unattended. It is a faster autocomplete, not a system — and per the timeline above, it is the half a model release replaces.

Directing an agent: five learnable capabilities

Directing an agent means owning the loop the agent runs inside, not just the prompt that enters it. Concretely it is five distinct capabilities, each learnable in order rather than absorbed by osmosis.

First, the loop itself: perceive state, plan, act, observe, decide whether to continue — the structure that turns one model call into a process that finishes a task. This is Agent Basics and Design Patterns.

Second, real tools and sub-agents: an agent that can only talk is a chatbot; one that can call your systems, delegate to a sub-agent, and compose results is a worker. Wiring that cleanly through something like MCP rather than ad-hoc glue is MCP & Skills and Multi-Agent Systems.

Third, designing for non-determinism: the same input does not produce the same output, so the engineering question shifts from "is it correct" to "what is the distribution of outcomes and what is the blast radius of the bad tail." A five-hour-horizon agent acting on real systems unattended — exactly the thing on OpenAI's timeline — is what makes this a design constraint rather than a footnote.

Fourth, evaluation before shipping: when the output is a paragraph, pass/fail tests stop working and "it looked good in the demo" is not a quality bar. Knowing how to measure a probabilistic system is Evaluation.

Fifth, running it in production: retries, fallbacks, cost per task, and the behavior when a tool call fails at 3 AM with no human watching. This is Production, and it is where most agent projects quietly die.

Why this is the actionable half

The salary numbers are real and inert — you cannot act on them. The paper treadmill is real and a trap — running it faster does not move the destination. The five capabilities are the opposite of both: each is something you can be unable to do on Monday and able to do later, in a defined order, on a real project. That asymmetry is the entire reason the distinction is worth extracting from an otherwise anxious thread. The salary is the headline; the loop underneath it is the work, and the only one of the three you can start this week.

Where to start

If you only use agents today, the gap is not a new prompt — it is the loop, the tools, the non-determinism, the evals, the production surface, roughly in that order. Start at Agent Basics and follow the loop outward; the rest is sequence, not mystery. Our companion piece reaches the same conclusion from the macro labor data rather than a campus thread (the redistribution, not the apocalypse). The thread's own best line is its disclaimer — the people actually at the top do not have time to post. The corollary is the point: they spent that time building, and what they built is on a syllabus.