# WP42 — Structured Literature Review

Working Paper 42 (UQO): "The Autonomous White-Collar Worker: Measuring the Economic Frontier of Claude Code, Computer-Using Agents, and Autonomous Knowledge Work"

Prepared by the literature agent, 2026-11-02. Every bracketed key [@key] resolves to an entry in literature/references.bib; every entry was verified against Crossref, the arXiv abstract page, or the official publisher page (see literature/VERIFICATION_LOG.csv). Papers that were sought but could not be verified are listed at the end and must not be cited.

Scope note: the bibliography deliberately prefers (i) peer-reviewed or NBER/IMF/ILO-series economics, (ii) arXiv preprints for agent benchmarks where the preprint is the canonical citation, and (iii) a very small number of industry documents (Anthropic computer-use announcement and Opus 4 system card, Stanford Digital Economy Lab working paper) whose publisher pages were fetched. Where an arXiv preprint has a later conference version, the review cites the arXiv record because that is what was verified; the writer should not add venue claims that were not checked.

# (a) Task-based models of automation and occupational exposure

The task framework is the backbone of the economics of automation and the natural home for the Autonomous Labor Frontier. [@autor2003skill] introduced the distinction between routine and non-routine tasks and showed that computer capital substitutes for routine cognitive and manual tasks while complementing non-routine analytic and interactive ones; the unit of analysis is the task, not the occupation, which is precisely the distinction the WP42 mandate asks us to preserve. [@acemoglu2011skills] generalised this into a Ricardian assignment model in which technologies and skill groups compete for tasks, and [@autor2015jobs] used it to explain why automation has historically created as many jobs as it destroyed. [@acemoglu2018race] and [@acemoglu2019automation] formalised the displacement, productivity and reinstatement effects: automation extends the set of tasks that capital performs, and the labour-market consequences depend on whether new labour-intensive tasks are created. Their empirical work on industrial robots [@acemoglu2020robots] remains the template for estimating local labour-market effects of a task-replacing technology. [@acemoglu2025macroeconomics] translates the same framework into an explicit accounting formula for generative AI: aggregate productivity gains equal the share of tasks exposed, times the share actually automated, times the average cost saving per task — an equation that depends on exactly the task-level success, cost and quality parameters WP42 proposes to measure experimentally rather than assume. [@restrepo2025missed] pushes the framework to its limit (AGI), distinguishing "bottleneck" from "supplementary" work; [@ide2025knowledge] models AI as a worker in a knowledge hierarchy that changes who solves which problems inside firms.

Occupational exposure indices are the empirical descendants of this framework. [@frey2017future] assigned a probability of computerisation to 702 occupations from expert labels and engineering bottlenecks; [@arntz2017revisiting] showed that using task-level heterogeneity within occupations cuts the share of jobs "at risk" from 38 % to about 9 % for the United States, a result that motivates measuring tasks rather than occupations. [@brynjolfsson2017workforce] and [@brynjolfsson2018machines] built the "Suitability for Machine Learning" rubric and found that most occupations contain some SML tasks but almost none are fully exposed, so reorganisation rather than wholesale replacement is the expected margin. [@felten2021exposure] linked AI capability benchmarks to abilities (AIOE), and [@felten2023language] re-weighted the index toward language modelling. The most-cited generative-AI exposure study, [@eloundou2024gpts], had humans and GPT-4 rate O*NET tasks for whether an LLM (alone or with software built on it) could halve task time at equal quality; roughly 80 % of workers have at least 10 % of tasks exposed and 19 % have at least 50 %. International institutions replicated the approach at global scale: [@gmyrek2023generative] and its refined 2025 index [@gmyrek2025refined] (ILO), and [@cazzaniga2024genai] and [@pizzinelli2023exposure] (IMF), which add a "complementarity" dimension so that exposed tasks are split between those likely to be automated and those likely to be augmented. [@hampole2025labor] construct firm-level, time-varying task exposure from job postings and find muted net employment effects because workers shift effort toward unexposed tasks.

All of these indices share a limitation that is central to WP42: exposure is an ex-ante judgement about what a model could do, not a measurement of what an autonomous agent does on real deliverables. Two newer strands partially close the gap. First, revealed-usage data: the Anthropic Economic Index [@handa2025economic] maps millions of Claude conversations to O*NET tasks and finds usage concentrated in software and writing tasks, with a majority of conversations classified as augmentation rather than automation; the September 2025 follow-up [@appel2025uneven] documents a rise in "directive" task delegation and heterogeneous enterprise API adoption. Microsoft's analogue [@tomlinson2025working] computes an "AI applicability" score from Copilot conversations. Early labour-market evidence remains mixed: [@brynjolfsson2025canaries] find employment declines for 22–25-year-olds in the most exposed occupations using payroll data, while [@humlum2025still], linking Danish adoption surveys to registers, find rapid adoption but small earnings and hours effects so far. Second, capability measurement on economically valuable deliverables — the subject of stream (c) below — is the direct precursor of WP42's design.

# (b) Field and laboratory experiments on generative-AI productivity

The randomised evidence on assistive generative AI is now substantial and remarkably consistent on the direction of first-order effects. [@noy2023experimental] randomised ChatGPT access among college-educated professionals doing mid-level writing tasks: time fell 40 %, rated quality rose 18 %, and the gains were largest for initially weaker writers. [@brynjolfsson2025generative] studied the staggered rollout of a generative-AI assistant to 5 000 customer-support agents and found a 15 % average increase in issues resolved per hour, concentrated among less experienced agents, with improved customer sentiment and lower attrition. [@peng2023copilot] randomised GitHub Copilot in a controlled programming task and found 55.8 % faster completion; [@cui2024developers] aggregate three field experiments inside large firms and estimate a 26 % increase in completed tasks. [@dellacqua2023jagged] gave BCG consultants GPT-4 and documented a "jagged frontier": inside the frontier consultants were faster and produced higher-rated work, but on a task deliberately chosen to lie outside the frontier AI users were 19 percentage points less likely to be correct. [@dellacqua2025cybernetic] extends this to teams at Procter & Gamble, finding that an individual with AI matched a two-person team without it. [@otis2024entrepreneurial] found that an AI business mentor in Kenya helped high-performing entrepreneurs and hurt low performers, a warning that heterogeneity by baseline skill can flip sign. [@choi2023lawyering] report a within-law-school experiment on legal drafting tasks with GPT-4, finding speed gains with uneven quality effects. [@vaccaro2024combinations] meta-analyse 106 experiments on human–AI combinations and find that, on average, the combination performs worse than the better of human or AI alone, except in content-creation tasks — a result that bears directly on how much human oversight of an agent adds or subtracts.

Two findings in this stream are warnings for WP42's design. [@becker2025metr] (METR) ran a randomised trial with experienced open-source developers using early-2025 AI tools on their own repositories: developers expected a 24 % speed-up and believed ex post they had been 20 % faster, but were measured to be 19 % slower. This is the only RCT in which measured and perceived effects diverge in sign, and it shows that self-reported productivity cannot substitute for timed, verified outcomes. [@agarwal2023radiology] show, in a radiology experiment, that giving humans AI predictions does not improve diagnostic accuracy on average because humans mis-weight the AI signal; the optimal design is task allocation rather than side-by-side assistance. Together with the decision-authority model of [@athey2020allocation] and the algorithm-aversion evidence of [@dietvorst2015aversion], these papers justify WP42's decision to measure autonomous agent performance separately from human-plus-AI performance (treatments A0–A4 versus H0–H1) rather than conflating the two.

What this stream lacks — and what WP42 targets — is any randomised evidence on fully autonomous, multi-application agents performing complete professional deliverables. Every study above evaluates a human using a chat assistant or an inline code completer; none measures the completion rate, cost, or supervision burden of an agent that owns the task end to end.

# (c) Agent benchmarks and computer-use evaluations

Benchmarks for autonomous agents have moved rapidly from single-function code generation to multi-application, long-horizon work. In software engineering, [@jimenez2024swebench] (SWE-bench) poses real GitHub issues with hidden unit tests; agent scaffolds such as [@yang2024sweagent] (SWE-agent) and the open platform [@wang2024openhands] (OpenHands) defined the agent–computer interface paradigm (file viewer, editor, shell) that Claude Code also instantiates. [@merrill2026terminalbench] (Terminal-Bench) extends this to hard, expert-written tasks in command-line environments and audits itself with the Agentic Benchmark Checklist of [@zhu2025rigorous], which catalogues outcome-validity and task-validity failures in existing agent benchmarks. [@miserendino2025swelancer] is particularly relevant economically: it prices 1 488 real Upwork freelance software tasks worth 1 million USD and measures how many dollars an agent "earns", with the best model at the time earning around 400 000 USD; [@chan2024mlebench] (MLE-bench) and [@wijk2024rebench] (RE-Bench, METR) compare agents to human experts on machine-learning engineering and research-engineering tasks under explicit time budgets, and find agents ahead at short budgets but humans ahead at long ones.

General computer use is measured by [@xie2024osworld] (OSWorld: 369 tasks in real Ubuntu/Windows/macOS desktops with execution-based checks, where early models solved about 12 % of tasks versus 72 % for humans), extended to long-horizon tasks in [@yuan2026osworld2]; by [@zhou2024webarena] (WebArena: self-hosted e-commerce, forum, CMS and map sites); by [@liu2024agentbench] (eight environments including OS, database and web shopping); and by [@mialon2023gaia] (GAIA: questions that are easy for humans and hard for assistants because they require tool use and multi-step browsing). Enterprise-flavoured benchmarks are the closest analogues to WP42's virtual white-collar economy: [@drouin2024workarena] (WorkArena, ServiceNow knowledge-work UI tasks), [@xu2024agentcompany] (TheAgentCompany: a simulated software company with Git, chat, wiki and co-worker agents, where the best agent completed about a quarter of tasks end to end), [@yao2024taubench] (τ-bench: retail and airline customer-service tasks with policy constraints and the pass^k reliability metric, which falls sharply with k), [@trivedi2024appworld] (AppWorld: nine simulated apps and 750 day-to-day tasks with state-based evaluation), [@wang2024officebench] (OfficeBench: cross-application office workflows), [@ma2024spreadsheetbench] (real-world spreadsheet manipulation), [@guha2023legalbench] (legal reasoning tasks built with lawyers), and [@backlund2025vendingbench] (Vending-Bench: a months-long simulated business whose long-horizon coherence collapses unpredictably). Anthropic's computer-use capability was introduced in [@anthropic2024computeruse] and its agentic safety evaluations (including prompt-injection resistance in computer use and "impossible task" behaviour in Claude Code) are documented in [@anthropic2025opus4card]; [@liu2026claudecode] offers the first academic dissection of Claude Code's design space (tool loop, context management, permission model).

The two benchmarks that come closest to WP42's economic question are GDPval and METR's time-horizon work. [@patwardhan2025gdpval] (OpenAI) builds 1 320 tasks from the work of professionals with 14 years' average experience across 44 occupations in the nine largest GDP sectors, with real files as inputs and deliverables (documents, spreadsheets, slides, diagrams) graded blind by experts against the human product; frontier models win or tie roughly half of pairwise comparisons, and the paper reports speed and cost multiples relative to experts. [@kwa2025horizon] (METR) defines the "50 % task-completion time horizon" — the length of task, measured in human-expert time, that a model completes with 50 % probability — and finds it doubling roughly every seven months on software tasks drawn from [@rein2025hcast] (HCAST) and RE-Bench. [@kapoor2024matter] argue that agent evaluations must report cost alongside accuracy and must hold out tasks to avoid scaffold overfitting; the same validity concerns are addressed systematically by [@zhu2025rigorous].

The gap WP42 occupies in this stream is specific. GDPval evaluates single deliverables produced from a prompt and files, graded against a human reference, without measuring the agent's interaction with a live computer environment, the number of dependent operations, the human interventions required, or the probability of recovery from errors. OSWorld, WorkArena and TheAgentCompany measure live computer use but in software or IT-service settings and without occupational weighting or professional-standard grading. METR's horizon is one-dimensional (task length) and software-only. None of them randomises tool access (shell vs. GUI vs. browser) on the same tasks, and none runs human baselines under the same protocol except in small calibration samples. WP42's A0–A4 design and 20-occupation benchmark would be the first to combine these elements.

# (d) Economics of AI agents, delegation and supervision

Delegating work to an agent that cannot be perfectly observed is the canonical principal–agent problem. [@holmstrom1979moral] shows that when effort is unobservable, any informative signal about it should enter the contract; [@holmstrom1991multitask] shows that when some dimensions of output are easier to measure than others, high-powered incentives distort effort toward the measurable dimension — an exact description of an agent that passes deterministic checks while neglecting unmeasured quality. These results motivate WP42's separation between deterministic (Agent E) and blind professional (Agent F) evaluation, and its explicit measurement of "critical errors" that a pass/fail test would miss.

The economics of AI literature re-frames these ideas. [@agrawal2019prediction] model AI as a fall in the cost of prediction, so that the complementary input — judgement about payoffs — becomes the scarce human contribution; [@athey2020allocation] model the choice of whether a human or an AI holds decision authority and show that AI can be optimally made less accurate so that the human remains engaged. [@korinek2023generative] catalogues the use of LLMs in economic research itself and introduces a useful taxonomy of delegation with verification. [@horton2023silicus] shows that LLMs can be used as simulated economic agents in experiments — a method WP42 must not use for its human baselines but may use for synthetic counterparties in the virtual economy. [@immorlica2024agents] ask what changes when generative AI acts as an economic agent in markets; [@hadfield2025economy] sketch the institutions an "economy of AI agents" would require; [@kolt2025governing] applies agency law to AI agents and argues that monitoring and loyalty are the central governance problems. [@sreenivas2026contracting] write down a contracting model for LLM delegation in which the principal chooses a verification cost and the agent chooses model and effort, deriving the agency cost of imperfect verification. [@catalini2026agi] ("Some Simple Economics of AGI") argue that as execution cost collapses the binding constraint becomes human verification, defining a "measurability gap" between what AI can do and what humans can economically check. This is the nearest verified precedent to the phrase "economics of verification" in the WP42 brief; no paper with that exact title could be verified.

Three further strands feed the supervision construct. In AI alignment, [@bowman2022oversight] define "scalable oversight" experimentally — can a human supervising a model outperform both alone? — which is the alignment-community counterpart of the supervision question WP42 poses in economic terms. [@svanberg2024costeffective] provide the clearest cost-effectiveness template: for computer-vision tasks they compute the share of exposed wage value whose automation is cost-effective at current prices (about 23 % of vision-exposed wages) and show how that share moves with system cost, which is exactly the budget parameter B in the ALF definition. Finally, the empirical fact that users increasingly delegate complete tasks (directive requests rising from 27 % to 39 % of Claude.ai conversations in eight months, [@appel2025uneven]) shows the supervision question is no longer hypothetical.

# (e) Reliability, long-horizon failure, compounding error and LLM-as-judge validity

A consistent finding across benchmarks is that agent success falls steeply with task length and with the number of dependent steps. [@dziri2023faith] show that transformer accuracy on compositional tasks decays rapidly with the depth of the computation graph and that apparent competence reflects pattern matching on shallow sub-graphs. [@sinha2025horizon] isolate execution from planning and show that small per-step accuracy differences compound into large differences in the length of task a model can execute, and that models exhibit a "self-conditioning" effect in which earlier errors raise later error rates. [@kwa2025horizon] model the success–length relationship as a logistic curve; [@yao2024taubench] show that pass^k (success on all k independent trials) is far lower than pass@1, so that reliability rather than peak capability is the binding constraint for deployment; [@backlund2025vendingbench] show qualitatively similar collapses over very long horizons. [@cemri2025fail] build a taxonomy of fourteen failure modes in multi-agent LLM systems from more than 150 traces (specification problems, inter-agent misalignment, and task-verification failures), which gives WP42 a coding frame for its own failure analysis. [@stroebl2024flaws] show that resampling with an imperfect verifier has a hard ceiling: beyond a point, drawing more samples increases false positives rather than correct solutions — a result directly relevant to any "self-repairing" treatment whose checker is itself a model.

On self-repair, [@shinn2023reflexion] (Reflexion) and [@madaan2023selfrefine] (Self-Refine) show large gains from verbal self-reflection and iterative self-feedback on coding and reasoning tasks when an external signal (test results) is available; [@chen2023selfdebug] show that models can debug their own code by explaining execution results. But [@huang2023selfcorrect] show that without external feedback, "intrinsic" self-correction of reasoning does not improve and can worsen accuracy. The implication for WP42's A4 treatment is that recovery should be defined relative to the type of error signal available (deterministic test, environment error, or none), and that an Autonomous Recovery Index must distinguish detection from correction.

For evaluation validity, [@zheng2023judge] (MT-Bench and Chatbot Arena) show that strong LLM judges agree with humans about as often as humans agree with one another, but document position bias, verbosity bias and self-enhancement bias. [@wang2024fair] show that simply swapping the order of two candidate responses can reverse an LLM judge's verdict; [@panickssery2024selfpreference] show that LLM evaluators recognise and favour their own outputs; [@gu2024survey] surveys the field and its mitigations (swapping, rubrics, multiple judges, calibration against human labels). The Claude-based evaluators in WP42 (Agent F) therefore require blinding to condition, randomised presentation order, rubric anchoring, and calibration against independent human experts, with agreement reported using the statistics of stream (f).

# (f) Experimental methodology: preregistration, inference, measurement models, survival analysis, agreement, cost-effectiveness

Preregistration and pre-analysis plans are now expected in experimental economics. [@nosek2018preregistration] separate prediction from postdiction and describe registration workflows; [@olken2015preanalysis] discusses the trade-off between credibility and flexibility in economics; [@simmons2011falsepositive] quantify how analytic degrees of freedom inflate false-positive rates. With many outcomes, occupations and treatments, multiple-testing adjustment is required: [@list2019multiple] adapt step-down procedures to experimental economics. For design and estimation, [@athey2016econometrics] is the reference on randomisation-based inference, stratification and power in field experiments.

WP42's data are clustered (repeated runs within scenario, scenarios within occupation, agents within model version) and the number of clusters at the occupation level is small. [@cameron2015practitioner] is the standard guide to cluster-robust inference; [@cameron2008bootstrap] introduce the wild cluster bootstrap for few clusters; [@abadie2023clustering] clarify that clustering should follow the sampling or assignment design rather than mere correlation, which for WP42 means clustering at the level of scenario assignment. Mixed-effects models are the natural way to pool across occupations and scenarios: [@bates2015lme4] for frequentist estimation, with Bayesian hierarchical models an alternative when occupation-level variance components are weakly identified.

Benchmarks are tests, and psychometrics applies. [@vania2021irt] fit item-response models to NLP test sets to compare their informativeness; [@polo2024tinybenchmarks] show that IRT-based selection of about 100 items can estimate benchmark performance within 2 points, which is the basis for scenario-difficulty calibration and for scaling from the 10-scenario pilot; [@zhou2025lost] apply IRT to LLM benchmarks and find many items uninformative or mis-keyed. [@miller2024errorbars] provide the simplest correct treatment of uncertainty in model evaluations (question-level clustering, paired differences, power analysis), and should be followed for every success-rate comparison in WP42.

Completion times are right-censored when an agent hits a budget, timeout, or emergency stop; survival analysis is the appropriate framework. [@kaplan1958nonparametric] for non-parametric completion curves, [@cox1972regression] for proportional-hazards regression on treatment and scenario covariates, and [@fine1999competing] for competing risks (success versus critical failure versus censoring), which is the natural model for the Professional Reliability Frontier because a "critical error" is a competing terminal event, not missing data.

Inter-rater agreement among blind evaluators should be reported with chance-corrected statistics appropriate to ordinal rubric scores. [@krippendorff2004reliability] and [@hayes2007reliability] define and justify Krippendorff's alpha for multiple raters with missing data; [@gwet2008interrater] proposes AC1/AC2 as a remedy for the paradox of low kappa under high agreement, which is likely when most deliverables pass. Both should be reported, with bootstrap confidence intervals.

For cost-effectiveness, the economic-evaluation literature's incremental cost-effectiveness ratio has a direct analogue: the incremental cost per additional professionally-acceptable deliverable, with uncertainty summarised by bootstrapping the joint distribution of cost and effect. Within AI evaluation, [@kapoor2024matter] propose reporting the Pareto frontier of accuracy against dollar cost and controlling for cost when comparing agents; [@svanberg2024costeffective] show how to convert task-level feasibility into the share of wage value that is cost-effective to automate; [@patwardhan2025gdpval] report cost and speed multiples relative to experts. WP42 should present all headline results as frontiers over budget B and quality threshold q rather than as single numbers.

# Positioning of WP42

This section states, for each proposed construct, the closest verified precedents and what would actually be incremental. The mandate instructs us not to assert novelty where a precedent exists; the honest summary is that none of the six constructs is new as an idea; the contribution would be their joint, experimental, occupation-weighted measurement under randomised tool access, with real human baselines.

Autonomous Labor Frontier (ALF). The ALF formula is a task-weighted joint probability of validated success, cost within budget, and human intervention within a cap. Each component has a precedent. Task weighting by occupational importance comes from the exposure literature [@eloundou2024gpts; @felten2021exposure; @gmyrek2025refined; @cazzaniga2024genai], and the formula is structurally the same as the "share of tasks automated" term in [@acemoglu2025macroeconomics] and the task-threshold condition in [@acemoglu2018race]. The budget condition is the cost-effectiveness share of [@svanberg2024costeffective]. Measuring success on real professional deliverables graded by experts is GDPval [@patwardhan2025gdpval]; measuring success as a function of task size is METR's time horizon [@kwa2025horizon]; "automation versus augmentation" classification of revealed usage is [@handa2025economic]. We found no verified paper that uses the name "Autonomous Labor Frontier" or that estimates the joint distribution of success, cost and human intervention from experiments on live computer environments. The incremental contribution is therefore (i) the joint-probability definition with explicit thresholds (B, H, q) rather than three separate indices, (ii) estimation from randomised agent runs with independent validation rather than from expert ratings or usage logs, and (iii) explicit sensitivity analysis over the task weights. The paper should present ALF as a measurement framework that operationalises the Acemoglu–Restrepo task model and the Svanberg cost-effectiveness logic, not as a new theory.

Workflow Dependency Frontier. The decline of reliability with the number of dependent steps is well documented: compositional depth [@dziri2023faith], execution length and self-conditioning [@sinha2025horizon], task length in human-time [@kwa2025horizon], pass^k reliability [@yao2024taubench], long-horizon collapse [@backlund2025vendingbench], and multi-application tasks [@xu2024agentcompany; @yuan2026osworld2; @wang2024officebench]. The incremental step is modest but real: treating the number of dependent operations, the number of applications, and the number of decision points as designed experimental factors in realistic office workflows across many occupations (rather than as incidental properties of software tasks), and estimating the per-step hazard with survival models [@cox1972regression; @fine1999competing]. The writer should cite METR's horizon and Sinha et al. as the direct precedents and describe the WDF as a multi-dimensional extension.

Autonomy-Adjusted Productivity. Adjusting output for oversight and correction time has precedents in the human-plus-AI experiments that measure time and quality jointly [@noy2023experimental; @brynjolfsson2025generative; @dellacqua2023jagged], in the METR RCT showing that perceived and measured productivity diverge [@becker2025metr], in cost-controlled agent evaluation [@kapoor2024matter], and conceptually in the verification-cost models [@catalini2026agi; @sreenivas2026contracting]. What is incremental is measuring human correction time and intervention counts inside an autonomous-agent experiment (A0–A4), so that productivity is net of supervision rather than gross of it, and reporting it as a function of the supervision cap H. The term "autonomy-adjusted" should be presented as a label for this accounting, not as a new concept.

Computer-Use Tool Premium. Benchmarks already compare agents with and without particular tools or observation spaces (screenshot vs. accessibility tree in [@xie2024osworld]; API vs. GUI in [@trivedi2024appworld]; scaffolding and prompting ablations in [@patwardhan2025gdpval]; agent–computer interface design in [@yang2024sweagent]). The incremental contribution is causal identification: randomising tool access (text-only, shell, shell + AppleScript/GUI, full computer use) across the same scenarios and occupations and reporting the premium as a treatment effect with clustered inference [@abadie2023clustering; @cameron2008bootstrap]. No verified study randomises tool access across occupations in this way; nonetheless, the construct is an ablation, and should be described as one.

Autonomous Recovery Index. Self-repair and reflection are established [@shinn2023reflexion; @madaan2023selfrefine; @chen2023selfdebug], as are the limits of intrinsic self-correction [@huang2023selfcorrect] and of imperfect verifiers [@stroebl2024flaws]; failure taxonomies exist [@cemri2025fail]; and pass^k already captures run-to-run reliability [@yao2024taubench]. The incremental element is the decomposition of recovery into detection and correction probabilities, estimated on naturally occurring and injected errors in professional workflows, conditional on the type of error signal available, and the comparison of a checkpointed self-repairing treatment (A4) with the same tools without recovery (A3). This is a measurement contribution; the mechanism is not new.

Professional Reliability Frontier. Multi-objective reporting of accuracy against cost is recommended by [@kapoor2024matter] and practised by [@patwardhan2025gdpval] and [@miserendino2025swelancer]; benchmark-validity standards are set by [@zhu2025rigorous]; the role of critical errors in professional contexts is anticipated by the jagged-frontier result [@dellacqua2023jagged] and the multitask-incentive logic of [@holmstrom1991multitask]. The incremental contribution is adding a critical-error dimension graded against professional standards (an independent terminal event in a competing-risks model [@fine1999competing]) and tracing the frontier over quality thresholds q. We recommend presenting this as a reporting standard rather than as a theoretical construct.

Overall positioning statement for the manuscript. WP42's defensible claim is that it is the first experimental study to (1) run autonomous, computer-using agents on an occupation-weighted benchmark of complete professional deliverables across at least 20 white-collar occupations, (2) randomise tool access and self-repair capability on the same tasks, (3) measure success, cost, human intervention, critical errors and recovery jointly with independent deterministic and blind professional evaluation, and (4) include observed human baselines under the same protocol. The closest precedents, which must be cited and compared explicitly, are GDPval [@patwardhan2025gdpval], METR's time-horizon and developer RCT [@kwa2025horizon; @becker2025metr], TheAgentCompany [@xu2024agentcompany], OSWorld [@xie2024osworld], τ-bench [@yao2024taubench], SWE-Lancer [@miserendino2025swelancer], and the exposure and cost-effectiveness indices [@eloundou2024gpts; @svanberg2024costeffective; @acemoglu2025macroeconomics]. Whether any of the six named constructs survives as a distinct concept should be decided after the pilot, and the names should be dropped if the results do not justify them.

# Papers sought but NOT verified (do not cite)

OpenAI, "Introducing SWE-bench Verified" (2024) — official page returned HTTP 403 to both fetchers; cite [@jimenez2024swebench] and describe the Verified subset in prose without a citation, or re-attempt verification before submission.
OpenAI, "Computer-Using Agent / Operator" (2025) — HTTP 403; not citable from this base.
Arntz, Gregory & Zierahn (2016), OECD Working Paper 189 — Crossref record carries no author metadata and the OECD page is bot-protected; replaced by the authors' 2017 Economics Letters paper [@arntz2017revisiting], which makes the same point.
Lane, Williams & Broecke (2023), OECD AI surveys — same author-metadata problem; dropped.
OECD Employment Outlook 2023 — fetched (DOI 10.1787/08785bba-en) but has no personal authors; omitted from the bib to keep the verification rule strict. The IMF and ILO reports cover the same ground.
Webb (2020), "The impact of artificial intelligence on the labor market" (SSRN) — not verified; not included.
Agrawal, Gans & Goldfarb, "Prediction Machines" (book, 2018) — not in Crossref; use [@agrawal2019prediction] instead.
"AI and the economics of verification" — no paper with this title could be found; the nearest verified works are [@catalini2026agi] and [@sreenivas2026contracting].
Toner-Rodgers (2024), "Artificial intelligence, scientific discovery, and product innovation" — withdrawn by MIT; must not be cited.
Chowdhury et al. / OpenAI "SWE-bench Verified" and Anthropic "Project Vend" — not verified.
Saunders et al. (2022) "Self-critiquing models for assisting human evaluators" (arXiv 2206.05802), Casper et al. (2025) "The AI Agent Index" (arXiv 2502.01635) and Yao et al. (2022) "ReAct" (arXiv 2210.03629) — verified on arXiv (records in /tmp/wp42-lit/) but cut from the final 110 to respect the size limit; they can be added on request without further verification work.
Conference-venue attributions (e.g., SWE-bench at ICLR 2024, OSWorld at NeurIPS 2024) were not verified and are deliberately absent from the bib; cite as arXiv.

spb/wp42_uqo