AGENT INSTRUMENTS · INTERACTIVE

Agent Work Observatory

Four small instruments for work shared by people and agents.

History reviewed · 2026-10-25

01 / PEAK TASK CONCURRENCY

Parallel tracks

Not recordedtasks

Personal measurement

Interactive demo · not measured data

60 min3 progressing together

02 / DOCUMENTED DELIVERIES

Delivery tickets

6items

Reviewed · 2026-10-25

Selected history

Development · Merged

Agent Work Observatory

View outcome and evidence
Delivered work
A dedicated page presents four interactive instruments, with a compact homepage entry and keyboard and touch controls.
Process and revisions
The large panel became four compact instruments, with research and methods collapsed below. Development moved to the Mac mini, followed by desktop and mobile checks.
Public evidence
PR #48 PR #52

03 / HUMAN ATTENTION

Attention clock

Not recordedhours

Personal measurement

Interactive demo · not measured data

Selected stage

30 min

Define objectives, prepare context, and break down tasks.

Select a stage · rework is included in total time

04 / REWORK SHARE

Rework tape

Not recordedhours

Personal measurement

Interactive demo · not measured data

15 minrework within 90 total minutes

6 documented deliveries reviewed; concurrency and human time are unmeasured. Demo values are excluded from measurements.

Research, records, and methods6 sources · recording templates

Reviewed history

Reconstructed from relevant conversations and project history, then checked against public changes. Revisions serving one goal count once. This review covers selected blog work.

Merge status and user acceptance are separate. This count describes review coverage and cannot establish a productivity gain.

Reviewed: 2026-10-25 · Reporting timezone: America/Los_Angeles

  • Agent Work Observatory
    View outcome and evidence

    A dedicated page presents four interactive instruments, with a compact homepage entry and keyboard and touch controls.

    The large panel became four compact instruments, with research and methods collapsed below. Development moved to the Mac mini, followed by desktop and mobile checks.

  • Bilingual Reading and Publishing
    View outcome and evidence

    Chinese and English pages share templates and navigation, with language switching, paired links, and bilingual search. The observatory also has an English edition.

    Later revisions restored English margin notes, strengthened reading checks, and added English synchronization from Chinese sources while preserving manual edits.

  • Reading TOC Follow
    View outcome and evidence

    The side table of contents follows the current section and keeps the final entry comfortably in view. Chinese and English articles share this behavior.

    Native smooth scrolling was replaced with a per-frame movement limit to reduce jumps when frames are skipped. Manual interaction can interrupt the follow motion.

  • Original Paper Figures
    View outcome and evidence

    Paper figures are shared across the author page, publication list, and detail page, preserving their proportions and colors with original-size viewing.

    Figure crops retain their proportions without a fixed canvas or upscaling. Manually selected previews take priority, and unknown PDF page numbers remain unset.

  • Shared Content Dossiers
    View outcome and evidence

    Author projects, recent posts, archives, and publication lists use shared dossier components with existing images, text, sources, and expandable details.

    Revisions adjusted headings, image and text hierarchy, and spacing; corrected empty image slots and mobile wrapping; and retained existing material.

  • Comment Section Width
    View outcome and evidence

    The comment section uses the article's reading width, correcting the narrow desktop editor while retaining the existing mobile layout.

    The container's full width was restored while preserving its maximum measure and internal structure. Rendered dimensions were checked at five viewport sizes.

Only work summaries and public project evidence are published. Message gaps, tool waits, and session spans are not human working time; concurrency requires run start and end records.

Personal log

Delivery tickets now include historical work. Daily concurrency, human attention, and rework time remain unmeasured.

Public research

Studies measure different outcomes. Keep each study's conditions and scope in view.

  1. Self-reported survey

    Record speed and output value separately

    In a survey of 349 technical workers, median self-reported gains across three output-value questions were 1.4–2 times, while the median self-reported speed gain was 3 times.

    Scope and original source

    A convenience sample and self-reported counterfactuals describe participants’ judgments, not proven causal gains. Record perceptions, actual time and delivery value separately.

    Read the original study
  2. Multi-agent benchmark

    Parallel gains depend on tasks and collaboration structure

    Version 3 compares 260 configurations across 6 benchmarks. Relative to a single agent, reported performance changes range from +80.8% for decomposable financial reasoning to −70.0% for sequential planning.

    Scope and original source

    These percentages describe benchmark performance on different tasks. They cannot form a single productivity curve and do not measure changes in human work efficiency.

    Read the original study
  3. Measurement update

    Multiple agents per person change time measurement

    METR notes that concurrent agent use, doing other work while waiting, and participant and task selection bias make per-task duration a poor proxy for actual productivity.

    Scope and original source

    This update exposes measurement limitations. It does not establish how many agents one person can manage, nor should new experimental point estimates be treated as certain productivity gains.

    Read the original study
  4. Randomized field experiment

    Feeling faster can diverge from being faster

    16 developers familiar with large open-source projects worked on 246 real tasks. Access to the AI tools available at the time increased completion time by 19%.

    Scope and original source

    These results concern specific people, repositories and a particular period. They cannot be generalized to all current development tasks. Review and correction belong in human effort.

    Read the original study
  5. Randomized field experiment

    Real work shows gains, with variation

    Three field experiments covered 4,867 developers. The pooled estimate found 26.08% more completed tasks with AI assistants, with a standard error of 10.3%.

    Scope and original source

    The research primarily examines coding assistants. More tasks do not directly imply greater value per person, and the studies did not measure multi-agent benefits.

    Read the original study
  6. Capability benchmark

    The horizon of achievable tasks is growing

    METR measures task difficulty by the time human experts need to complete it. During the study period, the task horizon at a 50% success rate doubled approximately every seven months.

    Scope and original source

    This measures human task time. It cannot be directly converted into continuous agent runtime, agents managed per person or reliable delivery capacity.

    Read the original study

Initial collection: 2026-10-24. These sources do not establish a consistently measured industry-wide trend in agents per person.

Methods and templates

Independent tasks
Count distinct acceptance objectives. Sessions, retries, and subagents for one task share an ID. Peak concurrency is the observed number of tasks progressing at the same time.
Reliable delivery
Meet the agreed criteria and retain a verifiable result link. Reviewed negative research results also count as valid deliveries.
Human effort
Include setup, additional context, review, coordination, and rework. Count each interval of attention once; rework is part of total time.
Comparable conditions
Compare similar task types, scope, and quality criteria. Leave missing values blank; enter 0 only for measured zero. Retain failures, abandoned work, and changes in tools.
Open dataset layers and null-value contract
Dataset layerGranularityNull-value and methodological boundary
Daily observations and tasksPer calendar day and independent acceptance goalUnmeasured concurrency or minutes stay null, never 0
Reviewed delivery historyGrouped by delivery goal with public PR evidenceCovers selected public history, not productivity multipliers
External empirical studiesIndexed per primary paper or field reportPreserves separate scopes across benchmarks, RCTs, and surveys

Daily records contain peak concurrency, not task start and end times. The minute-by-minute tracks are interactive illustrations; verified records remain the source of personal measurements. Open dataset licensed under CC BY 4.0 Creative Commons Attribution 4.0.