Who manages the agents?

An anonymous essay on off-policy.com, published under the slug "don't go quietly into the AI night" and organized around one question: who manages the agents? Its argument is not anti-AI. It's pro-AI and anti-concentration, which puts it closer to elena-verna-job-2027 than to the refusal pieces in anti-llm-discourse. The author discloses near the end that their company builds the exact product the essay prescribes, so the piece doubles as a vision statement and a pitch.

The two futures

The frame is a fork. One path is AI as a deity built and controlled by a small "clergy" — the labs that build frontier systems, get privileged access, and decide which capabilities everyone else may use. The clergy warn of mass unemployment (Amodei on entry-level white-collar jobs, Altman on customer support, Suleyman on professional work) and promise abundance and redistribution will compensate, "by the grace of the machines and their controllers." In that world a tiny group mediates between machine intelligence and everyone else, and most people are recipients of decisions and products, not participants. The other path: billions of humans each learning to direct capable agents of their own.

The essay's evidence for concentration is a capability-restriction trail. Drug discovery capped for bioweapons risk, software capped for cyber risk, and, projected forward, math capped for cryptographic risk and generation capped for disinformation risk — with governments requiring the limits and labs supporting them, while selected partners keep access. The result the author names is a two-tiered system: frontier AI for a small group, constrained AI for everyone else. (The citation trail mixes real 2025-26 events with forward-dated ones the reader should treat as speculative — export-control shutdowns, government-gated model tiers.)

The gap is the real argument

The strongest section is empirical, not rhetorical. Software development is the leading indicator because developers got agents a year before other knowledge workers. Among them, a sub-1% of super users have reached 100x (the Bun Zig-to-Rust rewrite in 11 days, one engineer running up to 64 concurrent Claude Code instances, is the showcase), while the median developer hasn't moved — the essay cites METR's randomized trial, where experienced developers were 19% slower with AI in early 2025 and a Feb 2026 follow-up found only "very weak evidence" of speedups. Every model and feature release widens the gap: the super user unlocks a level, the median gets more confused.

From this the author draws a distribution claim that is actually its thesis: a handful of 10x outliers barely moves the average; the way to move the average is to make the median person 2x. A company whose only AI strategy is headcount reduction "is using an exponential technology for a linear goal." This is the same distributional worry as llm-as-average-democratizer, pointed in the opposite direction — where that page asks whether AI raises the floor, this one insists the current trajectory raises only the ceiling.

The prescription: everyone becomes an agent manager

The build advice is to design for the median user, not the power user, so the user gains judgment and control — "design for the person managing the agent, not for the agent replacing the person." The organizational advice: stop selling (and stop buying) team-replacement. The people best placed to manage agents are the ones currently doing the work, because they carry the domain knowledge, and the management skills that transfer are ordinary ones — communicate intent, review others' work, own the output. Give each person one agent, one-to-one, then let good managers scale to a team and then a department of agents.

This is where the piece intersects the wiki's oversight cluster. Its "become an agent manager" ladder is the optimistic mirror of the skill-atrophy-supervision-paradox — it assumes the supervision skills exist and just need pointing at agents, where the paradox observes that heavy agent use erodes the very reading-and-critique skills supervision needs. It also lands near agent-principal-agent-problem: Crawshaw's point that agents collapsed the effort signal between contributor and reviewer is exactly the accountability hole this essay's "owned, controlled, accountable" refrain tries to close by naming a human owner for every agent. And it is the deliberate counter to ceo-ai-psychosis — where executive agent-orchestration simulates management without producing output, this piece wants management pushed down to the ICs who can actually verify the work.

Sovereign agents (and the pitch)

The closing move: agents should be "sovereign." Your company owns and controls their identities, permissions, memory, skills, artifacts, and audit trails; those assets are portable, governable, and inaccessible to anyone you haven't authorized. The load-bearing warning is "if it's not your team managing the agents, then your vendor is" — accepting a vendor-run agent hands out the access and data that automate your company away. The distinction between a human-owned agent and a vendor-run one echoes agents-vs-daemons: what the essay wants is agents with a clear human owner in the loop, not unattended processes running on someone else's chain of command.

What the HN thread pushed back on

The discussion surfaced the tensions the essay doesn't resolve. Timwi flagged the contradiction between "AI for billions" and agents "inaccessible to anyone you have not authorized" — it assumes 8 billion people each running a company still competing in capitalism; Terr_ answered that this is just the personal-computer democratization pitch (yours to own, not open to strangers). kristianc read "who manages the agents" as an enterprise-security anxiety surface in the lineage of "Shadow IT," the kind that always resolves into a priesthood and a six-figure governance platform — sharp, given the author sells that platform. ekidd noted the author's vision is itself unstable if models get much smarter: you can't hand everyone a personal SkyNet, and managing something smarter than you contributes little, so human management gets displaced too.

The most durable thread was accountability. simonw's proposal is to adopt the DRI (Directly Responsible Individual) and reserve it for a human who then uses all the agents they like — you can't put a ball of matrix arithmetic on a Performance Improvement Plan. Others sharpened the reason: a human has time preference and a body that can be fired or jailed, and that exposure is what makes accountability real. That's the mechanism the essay gestures at with "owned, controlled, accountable" but never quite states. Running underneath, several commenters supplied the reality check — current agents still get calendar scheduling wrong most of the time — against the essay's civilizational framing.