Human Attention as the Agent-Output Bottleneck

Sources#

• Cloudflare Cut 20%. Atlassian Cut 10%, Mostly Engineers. Bending Spoons Cuts Up to 75%. All Three Are Growing Faster or Earning More. So How Lean Can B2B Teams Really Get?
• How employment is changing at Anthropic
• The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers

Summary#

Most of the corpus's orchestration evidence counts agent effort: concurrent agents per person, agent-hours per day, agent-workdays per human workday (Parallel Agent Orchestration, AI Accelerating AI Development). Those series cannot say what the human is doing while the agents run, which is the open question several pages carry. Wilmers, Catipon, Massenkoff & McCrory, How employment is changing at Anthropic (Anthropic Institute Economics team, 2026-10-16) is the first source in the vault that measures the human side of the same telemetry inside one firm. Its answer is that the human becomes the slow step, and the work that grows is the work of being a less slow step.

Four measurements carry it:

• Agent hours outrun human hours. Counting only employee-directed interactive sessions (subagents pooled), agent runtime per human work-hour rises from ~0.12 in February 2026 to 6× by August. In research organizations it rises from ~0.2 to 19× (Figure 2, log scale, endpoints labelled). Interactive agents cross parity around early May (research orgs: early April). Adding fleet and autonomous agents, the total crossed human hours in March.
• The session is mostly waiting. Subtracting active agent time from each interactive session's wall-clock time, Claude sessions spend around 80% idly waiting for human input. "People are bottlenecking the agents."
• The output exceeds the reader. Converting interactive output tokens to reading time (0.75 words per token, 250 words per minute), reading everything would take 424 hours per person per week in engineering and research, 43 in finance and accounting, 35 in sales, 27 in other non-engineering, 26 in legal, policy and trust, and 12 in marketing (Figure 6). Most non-technical groups are past half a 40-hour week.
• Meta-work doubles. An automated coder over 2,000 one-line session summaries finds the share of sessions spent on building skills, plans and reviews for other agents, dashboards and monitors for agent pipelines, and documenting how work is done rising from ~10% (Aug–Oct 2025) to 22% (Apr–Jun 2026) for the average employee (Figure 7). Most of the rise is agent management specifically: specs for multi-agent runs, looping one agent to review another, monitoring pipelines.

Evidence note. empirical (tier corrected from the raw's case-study; reasoning in Source Notes), with three limits that travel with every number. (1) First party. Anthropic's economists measure Anthropic's employees using Anthropic's product, and the report is published by the company that sells the agents. (2) One firm, at the frontier. The report itself calls Anthropic a firm that "has thoroughly molded its workflows around AI". It is a preview, not a population. (3) Constructed denominators. Human hours in the 6× ratio are headcount × an assumed 40-hour week, not measured time. Agent time is runtime including inference and active tool use but excluding any wait longer than 30 minutes. The reading load is a token conversion, not a measure of what anyone read, and both the meta-work share and the hiring-scorecard shifts come from automated coders whose agreement with humans is not reported. Figure 6 and 7 start-of-period values are read off the plot and approximate. Endpoints are labelled.

The reading load is a ceiling, not a workload#

Figure 6's 424 hours a week is not a claim that engineers read for 424 hours. It is the volume that cannot be read. Something other than reading has to decide what gets a human's attention, and that is where the meta-work goes. The report's examples are the instruments of selective attention: an agent "chief of staff" that triages other agents' requests, recruiters' dashboards in which agents surface problems and start actions, an engineer's "performance review" of their Claude team's communication style. Standing written instructions for agents grow from a few hundred in July 2025 to 8,500 a year later. That is the codify-once margin Agentic Work Systematization measures as skill adoption at OpenAI, read here as the volume of the instruction base rather than the share of users.

The function split matters. Meta-work rises most in infrastructure (~5% → 29%), developer tooling (~5% → 26%), security and trust (~14% → 27%), product engineering (~13% → 26%) and safety research (~11% → 25%). It stays flat in sales (~15% → 15%) and is lowest in finance, legal and marketing (~6% → 12%). The functions that run fleets build tooling to supervise them. The functions whose core work is a conversation with a customer do not have that option.

The emotional register is AI Brain Fry's. One engineer describes the job as facing "an army of Claudes, with every one asking for my attention", and another compares it to babysitting, where "every child brings to you concerns with an equal demand for your attention". In the April 2026 survey, 20% of open-ended answers described their role shifting from doing work to managing Claudes, against 27% describing productivity gains.

What stays human, and grows#

Meetings rise. Calendar data puts the median employee's meeting time 20% above a year earlier, and 16% higher for technical individual contributors, the subgroup least affected by the company's shift toward meeting-heavy non-technical roles. The interview reason given is that chasing a bug increasingly "uncovers these organizational or systemic or tech debt things that actually require some amount of human convening and decision-making and governance to solve." In the report's phrase, "a break from managing the agents typically looks like jumping into a meeting."

Agents add to human communication rather than replacing it. Agents appear in nearly half of active public Slack channels. In one large department's thread network (Figure 8, not captured in the raw), human threads keep the same number of human responses and layer agent responses on top. Agents cluster in dedicated Claude-human channels and connect to one employee each: they "become teams for a single employee, rather than nodes with high connectivity to many employees at once." So the attention load is per-person rather than shared. Each employee runs a private team whose output only they are positioned to read.

Human work has to become agent-visible. A decision made in an unlogged meeting is missing context for every agent working from the written record. The report's example is an employee whose name on a months-old document about Claude swag kept agents sending them t-shirt and mug requests long after the job had moved on. The report's conclusion is normative: "Private work, in unlogged meetings and direct messages, is invisible, confusing and discouraged in an AI-heavy workplace." This is the human-to-agent direction of the bottleneck. The person is a reader of agent output and also a writer of the context agents read, and an unwritten decision is stale context the moment it is made.

Against the rest of the corpus#

• The ratio, against OpenAI's. METR's Chris Painter quotes OpenAI's research organization at 3.1 agent-workdays per human workday in mid-August 2026 (AI Accelerating AI Development). Anthropic's research-org figure for the same month is 19× agent-hours per human hour. They do not compare. OpenAI's unit and session scope are not given, and Anthropic's ratio uses an assumed 40-hour human week and counts only employee-directed interactive sessions. What they share is Painter's reading: agent activity already outruns the humans who would read it.

• The idle share, against concurrency. Parallel Agent Orchestration's Codex telemetry shows 28.6% of OpenAI staff peaking at 5+ concurrent agents. An 80% idle share is what concurrency looks like from the session's side. If each session waits four-fifths of its life for one person, running several is how that person keeps their own time full, and each added session adds a queue the person must service.

• The cost, against the system view. Task-Level Speedup, System-Level Cost argues that AI's saving at the producer is booked as cost somewhere else in the organization. Here the "somewhere else" is the same person's attention, a day later: the speedup produces output the producer cannot read, and the meta-work is the price of not reading it.

• The review step. Verification as the New Bottleneck locates the bottleneck at review. This source finds something earlier than review and broader than code. The bottleneck is whether a human reads the output at all, and it applies to the 26–43 hours a week of non-technical output as much as to code. Within-person code-quality proxies show no AI-versus-human difference on the same engineer's code (see AI Accelerating AI Development), so the attention problem here is about volume, not about AI code being worse.

• The communication finding, against a controlled design. Generative AI and Collaboration Networks (Soda, Iorio & Rizzo, empirical) finds the same direction with a control group: across 11,868 knowledge workers at a consumer-products multinational, Copilot adoption raised the hours workers convene with colleagues (+1.43 h/month over three months, mostly meetings and calls) and widened their networks. That is an assistant rollout, not an agent fleet, and its proposed mechanism is validating AI output rather than convening on systemic problems. Two firms, two instruments, and neither finds displacement.

• The same shift at three people. Function-Selective AI-Era Headcount Cuts carries Lemkin's account of SaaStr (3 humans, ~20 agents, self-reported, practitioner-opinion): a 13-hour day spent "checking, re-ranking, and fixing, with no new features built", and the rule "someone has to own what the agents produce". It is an anecdote, not a measurement, but it is the doing-to-reviewing shift this page measures, at a firm where the owner is also the only reviewer.

Limits#

• No outcome for the attention. Nothing measures whether the idle time, reading load or meta-work cost anything: no error rate, no fatigue measure, no time-use diary. The bottleneck is inferred from agent-side telemetry plus interviews.
• Meta-work is classified, not timed. The 22% is a share of sessions labelled meta-work by an automated coder over a 2,000-summary sample, not a share of hours. A session that writes a spec and a session that writes code count the same.
• Composition. The firm tripled in size in a year and shifted toward non-technical roles, which carries part of the meeting rise (the report says so). The technical-IC cut (+16%) is the cleanest figure, and it is still not adjusted for team size or tenure.

Connections#

• Parallel Agent Orchestration — the agent-side counterpart: concurrency and agent-hour telemetry, which this page reads from the human's side (idle share, reading load)
• Founder as Agent Orchestrator — the decision-burden question there asks for a measurement of oversight load. This source supplies idle and reading-volume telemetry for employees (not founders), with no direct measure of attention
• AI Brain Fry — the cost this page's numbers imply and do not measure: fatigue from oversight beyond capacity, here described in interviews as babysitting an army of Claudes
• Agentic Work Systematization — the codify-once response at OpenAI, measured as skill adoption. Here it is measured as standing-instruction volume (hundreds → 8,500 in a year) and as a meta-work share of sessions
• AI-Native Organization — Tan's "everyone is a manager of agents now", measured at one firm: 20% of survey answers describe the shift from doing to managing, and meta-work reaches 22% of sessions
• AI Accelerating AI Development — the output side of the same firm's telemetry (commits, agent hours, task difficulty after model releases) and Painter's oversight reading of OpenAI's 3.1 agent-workdays
• Task-Level Speedup, System-Level Cost — the producer's saving booked elsewhere. Here the cost returns to the producer's own attention as unread output and as tooling to filter it
• Agentic Coding Work-Composition Shift — the same lab classifying sessions by purpose. That study's work modes move from fixing toward operating and analysis; this one adds a mode the nine-mode taxonomy has no bucket for, which is work about the agents themselves
• Systems Thinking Over Specialization — the hiring consequence the same report measures: interviewers credit autonomy and judgment, and discuss domain expertise less
• Seniority-Biased AI Adoption: The Junior Share at Adopting Firms — the employment consequence measured in the same report: Anthropic's junior technical share falling while agent hours rise
• Verification as the New Bottleneck — the review-step version of this bottleneck. This page's version sits upstream of review and outside code
• Function-Selective AI-Era Headcount Cuts — the headcount consequence: a cut has to be sized by how much agent output the remaining humans can review (Lemkin: "someone has to own what the agents produce")
• Anthropic / Anthropic Institute — the firm studied and the team that studied it
• Generative AI and Collaboration Networks — the same no-displacement finding under a staggered DiD with never-adopter controls (Copilot, 11,868 workers): convened hours up, networks broader and less clustered, more time on the strongest prior ties

Open Questions#

• Is the ~80% idle share a property of Anthropic or of interactive agent sessions generally? Settler: session telemetry from another firm or a vendor's population that computes wall-clock time minus active agent time on the same definition, ideally split by concurrent-session count.
• How much of the unread output is never read by anyone? Figure 6 is a ceiling built from token counts. Settler: telemetry of human read or review actions per output token (scroll, open, approve, edit), so that reading coverage can be set against output volume by function.

Sources#

• Cloudflare Cut 20%. Atlassian Cut 10%, Mostly Engineers. Bending Spoons Cuts Up to 75%. All Three Are Growing Faster or Earning More. So How Lean Can B2B Teams Really Get? — Jason Lemkin, SaaStr, 2026-10-17 (practitioner-opinion). Cited only for the SaaStr review-load anecdote and the "someone has to own" rule, both self-reported
• How employment is changing at Anthropic — Nathan Wilmers, Regina Catipon, Maxim Massenkoff & Peter McCrory (Anthropic Institute, Economics), How employment is changing at Anthropic, anthropic.com, 2026-10-16, empirical (tier corrected from case-study). Internal employment, agent-session telemetry, output, survey, Slack, calendar and hiring-scorecard data plus 17 interviews. Cited here for Figure 2 (agent hours per human work-hour, viewed), the idle-time and reading-time computations, Figure 6 (reading hours by function), Figure 7 (meta-work by function, start values read off the plot), the standing-instructions count, the April 2026 survey's open-ended shares, calendar meeting time, the Slack network prose (Figure 8 itself was not captured) and the agent-visibility argument. COI: first-party. Anthropic studies its own staff using its own product and publishes the result
• The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers — Soda, Iorio & Rizzo (Bocconi), arXiv 2610.32461, 2026-10-23, empirical. Cited only for the controlled-design comparison in Against the rest of the corpus; full treatment on Generative AI and Collaboration Networks