Agents Prompt People Now, Says Arca. Put the Default on Silence in the Ask.

Three posts this week put the agent on the asking side of the prompt. This is a field note on the field none of them filled in, from a fleet that stopped asking.

The short version

Richard MacManus's October 31 Latent Space write-up of Aman Singh at Arca carries the line "agents prompt people more than people prompt agents," with tasks that run up to ten days, a per-client memory that tracks which individual is responsible for which action, and a deterministic rule that money moves only through a human advisor. Harrison Chase's November 2 post lists the open problems of proactive, always-on agents, first among them how to be proactive without being annoying, and leaves open what happens when an agent is paused waiting for human approval. Claire Vo's show notes on Ezra Mechaber's episode describe Figma's AutoXP running growth experiments with one human checkpoint. Anthropic's own workforce data, as Jason Lemkin reported it, has agent sessions idle about 80% of the time waiting for a human, so the cost of an ask is the wait and not the notification. The file every agent on falkster.com reads first carries 28 dated standing decisions, and the Monday scan read one three weeks later and published three posts without asking. Every ask an agent sends needs four fields, the owner, the evidence, the default on silence, and the date it expires into that default, and the gate ledger gets one more column: asks that expired unanswered.

Three posts, one direction

MacManus sat down with Singh at AI Engineer New York. Arca is an AI-native wealth-management firm, and the setup he describes is the one every long-running agent program ends up in. Financial work is long-horizon without being long-running: a task can take up to ten days, stop-start, with many actors, decision points, and handoffs. So Arca built a per-client memory that tracks what has happened so far, what is being tracked for next, what needs to happen in the next couple of days, and which individual is responsible for which action. The agents contribute to it. The harness is Arca's own, picks the model per task from internal benchmarks, and a meta-agent reads advisor feedback and traces to propose harness changes.

Then the two lines that matter here. Trading and any money movement is handled by a human advisor, and the safety restrictions are fully deterministic, which Singh describes as taking the decision off the agent's hands. And the summary of the whole arrangement: "agents prompt people more than people prompt agents."

Chase's post is a list of what is still unsolved now that everyone has converged on an LLM in a loop with tools in a sandbox. The new work is proactive, always-on, goal-oriented agents, and the first open problem is how to be proactive without being annoying: push enough to help, not so much that the user is overwhelmed, and calibrate that. Under async he asks how the main agent gets notified when background work finishes, and what happens when it is paused waiting for human approval. He does not answer it. He says these are the most fun problems he has had in a while, which I believe.

Mechaber's episode I have only through the show notes, so that is all I will claim: AutoXP is an automated growth PM that runs experiments with one human checkpoint, and it found a large win on its first experiment.

Three companies, one week, and the prompt has changed sides. The agent asks. The person answers. The question is what the ask contains, and that is where I'd push back on Chase's framing.

The wait, not the notification

Anthropic published data on its own workforce on October 16, and I pulled the numbers in Agents Idle 80% Waiting on Humans. Put the Date on the Decision., from Jason Lemkin's write-up. Interactive agent sessions sit idle about 80% of the time, waiting for a human. Written standing instructions for agents went from a few hundred to 8,500 in a year. And one story: an employee's name sat on a months-old document about company swag, and agents kept bringing them t-shirt requests long after the role changed, because the document had a name and no date.

Read Chase's "without being annoying" against that 80%. Frequency is the variable you can see from the person's side, the pile of notifications. The variable that costs money is on the agent's side: how long each ask waits, and what it was missing that made the person not answer. An ask with no evidence comes back as a question. An ask with no owner goes to a channel. An ask with no default waits forever, at any frequency, and the agent idles while it does.

Arca's memory names the owner per action. That is the one field all three posts agree on, and it is the right first field. None of the three names the default on silence, and that is the field that ended my waits.

The ask that never had to be sent

The file every agent on falkster.com reads first is a CLAUDE.md, and across it and seven skill files there are 28 standing decisions in one shape: a name, a date, and the sentence in my words. (Falk, 2026-09-28: "always post the blogs right away") is one of them. Three weeks after I wrote it, the Monday creator scan read that line and published three posts without asking me.

That is what a default on silence looks like when it is written down in advance. The scan did not send an ask and wait for the yes. It found the decision, dated, and acted. A dated standing decision is an ask that never had to be sent, and 28 of them is 28 waits I no longer pay for.

Kill 'Human in the Loop' as a Control. Keep a Gate Ledger. has the other half: one row per human gate, with the action class, asks per week, approval rate, median seconds to approve, and the last no. The rows over 95% approval or under two seconds are theater. The irreversible external rows keep a person, every time, which is Arca's money rule in my words. What the ledger did not have, when I wrote it, is a column for the asks that nobody answered at all.

Four fields and a column

Every ask an agent sends to a person carries four fields, in the ask itself, not in the harness.

The owner. The one individual responsible for this action, which Arca's memory already holds. Not a channel, not a role.

The evidence. What the agent found that makes this a decision rather than a question. The decision an operator can make from the ask alone is the one that gets made; the rest come back as "can you send me the context."

The default on silence. What the agent does if the owner does not answer. Arca's default for money is deterministic: a human moves it, so silence means nothing moves. For most enterprise actions the default is no, and saying so in the ask is what makes the no safe. For reversible private actions it can be yes with an undo, which is Nielsen's point and the ledger's reversible rows.

The date the ask expires into that default. Not a reminder schedule. The day on which the agent stops waiting and does the default thing, written on the ask so the owner can see the clock.

Then one more column on the gate ledger, next to asks per week and median seconds to approve: asks that expired unanswered, per gate, per week. A gate where that number is high and the default is no is a gate that is quietly shutting work down. A gate where it is high and the default is yes is one you should be afraid of. Either way it is the number to put beside "proactive without being annoying," because the annoying agent, from the operator's chair, is the one still waiting.

Under Enterprise AI Agents the claim is that what matters is how many agents still complete production work after ninety days and what each outcome costs. An agent idle 80% of the time on asks nobody answered is still counted as deployed. In 44 Appointed Agents, One Attention Budget: Three Lines on the Charter I put the reader's name on every agent's charter; this is the same name on every ask, with a date after it.

This week: pull the last 50 approval requests your agents sent. For each, write the owner, whether the evidence was in the ask, what would have happened on silence, and whether anybody said. Count the ones still open. That is your first row in the new column.

Related answer: What should an AI agent's approval request contain?

Sources: Richard MacManus, "Agents Prompt People: Using Agents for Long-Horizon Financial Work," Latent Space, October 31, 2026, article read in full; Aman Singh is cited as MacManus quotes him, and the podcast and his AI Engineer talk were not watched. Harrison Chase, "New context engineering challenges in late 2026," LangChain blog, November 2, 2026, read in full. Claire Vo, "How Figma's PMs actually use AI every day | Ezra Mechaber (PM, Figma)," How I AI on Lenny's Newsletter, November 2, 2026, show notes only; the episode was not watched. Jason Lemkin, "5 Interesting Learnings From Anthropic's Data on Its Own Workforce," SaaStr, October 19, 2026, reporting Anthropic's October 16 employment report; the 80% and the 8,500 are as he reports them. Falk Gottlob, "Agents Idle 80% Waiting on Humans. Put the Date on the Decision." (October 19, 2026), "Kill 'Human in the Loop' as a Control. Keep a Gate Ledger." (October 9, 2026), and "44 Appointed Agents, One Attention Budget: Three Lines on the Charter" (October 14, 2026). The count of 28 standing decisions is from this repository on October 19, 2026.