| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": {"@id": "cr:data", "@type": "@json"}, |
| "dataType": {"@id": "cr:dataType", "@type": "@vocab"}, |
| "dct": "http://purl.org/dc/terms/", |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "name": "AgentPulse", |
| "description": "AgentPulse is a continuous multi-signal evaluation framework for AI agents. It scores 50 agents across 10 workload categories along four factors (Benchmark Performance, Adoption Signals, Community Sentiment, Ecosystem Health), aggregated from 18 real-time signals collected from GitHub, package registries, IDE marketplaces, social platforms, and benchmark leaderboards.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "license": "https://creativecommons.org/licenses/by/4.0/", |
| "version": "v6", |
| "datePublished": "2026-05-04", |
| "creator": {"@type": "Organization", "name": "Anonymous (NeurIPS double-blind submission)"}, |
| "keywords": [ |
| "AI agents", "LLM evaluation", "benchmarking", |
| "deployment evaluation", "sentiment analysis", "software engineering" |
| ], |
| "citeAs": "@inproceedings{agentpulse2026, title={AgentPulse: A Continuous Multi-Signal Framework for Evaluating AI Agents in Deployment}, author={Anonymous}, booktitle={NeurIPS Datasets and Benchmarks Track}, year={2026}}", |
| "isLiveDataset": true, |
|
|
| "rai:dataCollection": "Signals were collected on independent schedules from public APIs and leaderboards: GitHub REST/GraphQL (stars, forks, contributors, commits, issues), PyPI BigQuery and npm registry (downloads), VS Code Marketplace public Gallery API (installs, ratings), and social platforms (Bluesky, Reddit, Hacker News, Stack Overflow, GitHub Discussions, Dev.to, Mastodon, V2EX). Benchmark scores were scraped from the official leaderboards of SWE-bench, GAIA, WebArena, HumanEval+, and tau-bench. Collection cadence ranges from 5 minutes (high-frequency social platforms) to 24 hours (benchmarks). All data was passed through a four-component data-quality protocol (uniqueness, bot detection, source credibility, specificity) before NLP scoring. The 50-agent registry was hand-curated based on prominent agentic systems with public visibility as of April 2026.", |
|
|
| "rai:dataLimitations": [ |
| "Snapshot from 2026-04-20 to 2026-05-04; data evolves continuously and any single snapshot is a point-in-time view.", |
| "Only 19 of 50 agents have published benchmark scores; the remaining 31 receive a neutral prior of 0.5 on the Benchmark factor, which compresses cross-agent variance on that dimension.", |
| "Sentiment data has narrow distribution across agents (range [0.002, 0.214], mean 0.06, std 0.07), limiting the Sentiment factor's discriminative power.", |
| "Stack Overflow question counts are tag-based and subject to alias noise (e.g., the 'cursor' tag includes both Cursor IDE and unrelated CSS-cursor questions). Mention-filtering applied at collection time partially mitigates this; raw tag counts in the released snapshot do not.", |
| "Closed-source agents (Devin, OpenAI Codex, Cursor, Operator, Manus, ChatGPT, Claude) lack observable signals on the Adoption factor by construction (no public repository, package, or marketplace listing); the released snapshot reports zero on those signals, which is a measurement boundary rather than a quality verdict.", |
| "The NLP sentiment pipeline is English-only; agents with strong adoption in non-English communities (V2EX, Chinese developer forums, etc.) receive attenuated sentiment signal.", |
| "Per-agent mention counts are heavy-tailed (median 24 texts, max 7,912); agents with low mention counts have noisier sentiment estimates.", |
| "The 50-agent registry mixes standalone agents (Claude Code, Cursor, Devin) with agent frameworks (LangGraph, LlamaIndex, CrewAI). Frameworks are not directly benchmarkable on agent-task suites such as SWE-bench, which is a structural limit on cross-population comparison." |
| ], |
|
|
| "rai:dataBiases": [ |
| "Selection bias: posters on social platforms are not representative of the broader user base of any agent. Power users and early adopters are over-represented; passive or enterprise users are under-represented.", |
| "Platform credibility weighting (Stack Overflow 0.90 vs Bluesky 0.60) introduces systematic differences in how each platform's sentiment contributes to the composite, reflecting moderation rigor rather than user authenticity.", |
| "Registry construction bias: the 50-agent set was hand-curated based on visible/popular agents in the agentic-AI ecosystem; less-publicized agents (research prototypes, internal corporate agents, regional or non-English-speaking ecosystems) are systematically excluded.", |
| "Open-source vs. closed-source asymmetry: GitHub-derived signals (stars, forks, contributors, issues) favor open-source agents. Closed-API agents have no observable adoption signal under the framework's measurement scope.", |
| "Benchmark availability bias: agents whose authors prioritize academic evaluation (e.g., publishing on SWE-bench Verified) receive measured Benchmark scores; agents that target consumer adoption without leaderboard submissions receive a 0.5 prior.", |
| "Temporal bias: newer agents have shorter observation windows and noisier per-agent estimates than long-established agents; the snapshot does not adjust for cohort age.", |
| "Language bias: the English-only NLP pipeline systematically attenuates the Sentiment factor for agents with significant non-English communities.", |
| "Confirmation bias risk: the framework's design weights (w_B=0.35, w_A=0.25, w_S=0.20, w_E=0.20) are a principled prior but not empirically optimal; reweighting changes rankings." |
| ], |
|
|
| "rai:personalSensitiveInformation": [ |
| "The released CSV bundle (data/csv/) contains no individual user records, no per-text data, and no PII. It contains only per-agent aggregates (counts, means, time-series).", |
| "The historical SQLite snapshots (data/sqlite/arena_april20_v0.db, arena_april20_v1.db) and the agent-subset SQLite (arena_agent_subset.db) similarly contain only per-agent aggregated signal data.", |
| "No real names, email addresses, demographic information, geographic location, or private contact details are included.", |
| "Author handles or post identifiers from social-media platforms are NOT included in any released file. Post-level data with public author handles is collected by the pipeline at runtime but is not shipped in this artifact.", |
| "Per-text content (raw social-media posts, GitHub issue text, Stack Overflow questions) is similarly NOT included in this release; only aggregated sentiment scores derived from that content are released.", |
| "All collected data originates from public APIs operating under each provider's terms of service; nothing in this release was obtained from private accounts, paywalled content, or non-public sources." |
| ], |
|
|
| "rai:dataUseCases": [ |
| "Research on multi-signal evaluation methodologies for AI agents and other rapidly-evolving software systems.", |
| "Comparative analysis of capability metrics (benchmark scores) versus adoption metrics (downloads, installs, community engagement) in the AI tool ecosystem.", |
| "Longitudinal studies of how AI-agent adoption evolves over time as the snapshot is refreshed via the released collection pipeline.", |
| "Methodological research on circularity-controlled validation of composite evaluation scores.", |
| "Sensitivity and robustness analyses (factor weight perturbation, leave-one-out, Dirichlet bootstrap) for multi-criteria decision frameworks.", |
| "Educational use as a worked example of a continuous-evaluation pipeline combining heterogeneous public APIs with NLP scoring.", |
| "Auditing and reproducibility studies of multi-source evaluation systems." |
| ], |
|
|
| "rai:dataSocialImpact": "AgentPulse rankings could shape developer and procurement decisions about which AI agents to adopt; users should treat any single snapshot as one signal among many rather than a definitive quality verdict. Two specific risks warrant explicit acknowledgment. First, the framework systematically under-credits closed-source agents that have no observable adoption signal under the current measurement scope (e.g., Devin, OpenAI Codex, Cursor, Operator); we recommend interpreting any AgentPulse ranking that involves closed-source agents alongside benchmark-only rankings rather than as a substitute. Second, sentiment-based scoring may amplify popular narratives or marketing-driven discourse over substantive technical quality, particularly for new releases with high social-media discussion. Downstream researchers and practitioners are encouraged to (i) reweight the four factors for their specific use case via the released code, (ii) consult the per-factor decomposition rather than only the composite, and (iii) cite the snapshot date when reporting AgentPulse-derived rankings.", |
|
|
| "rai:hasSyntheticData": false, |
|
|
| "rai:dataAnnotationProtocol": "No human annotation was performed for the released aggregates. Sentiment scores are computed by an automated four-model ensemble (VADER, TextBlob, FinBERT, DistilBERT-SST2) at collection time. A 200-text held-out manual annotation set was used internally for pipeline calibration (Cohen's kappa = 0.81 between two annotators); this calibration set is not part of the released bundle.", |
|
|
| "rai:dataReleaseMaintenance": "The dataset is released as a static snapshot (version v6, dated 2026-05-04). Collectors and scoring code are released alongside the data so that maintainers and downstream researchers can refresh signals against live APIs. The authors do not commit to a fixed release cadence beyond the camera-ready submission, but the released code is designed for continuous re-collection.", |
|
|
| "distribution": [ |
| {"@type": "cr:FileObject", "@id": "agent_scores.csv", |
| "name": "agent_scores.csv", "contentUrl": "data/csv/agent_scores.csv", |
| "encodingFormat": "text/csv", |
| "description": "Composite + per-factor scores per agent per category, time-stamped (13,504 rows). The headline scoring output."}, |
| {"@type": "cr:FileObject", "@id": "agent_signals_raw.csv", |
| "name": "agent_signals_raw.csv", "contentUrl": "data/csv/agent_signals_raw.csv", |
| "encodingFormat": "text/csv", |
| "description": "Per-agent aggregated 18-signal observations (3,283 rows). Includes github_stars, vscode_installs, sentiment_avg, bench_swebench, etc. as a JSON blob per row."}, |
| {"@type": "cr:FileObject", "@id": "agent_benchmark_signals.csv", |
| "name": "agent_benchmark_signals.csv", "contentUrl": "data/csv/agent_benchmark_signals.csv", |
| "encodingFormat": "text/csv", |
| "description": "Published benchmark scores: SWE-bench, GAIA, WebArena, HumanEval+, tau-bench (1,563 rows)."}, |
| {"@type": "cr:FileObject", "@id": "agent_github_history.csv", |
| "name": "agent_github_history.csv", "contentUrl": "data/csv/agent_github_history.csv", |
| "encodingFormat": "text/csv", |
| "description": "GitHub stars, contributors, and commits time-series per agent (62 rows)."}, |
| {"@type": "cr:FileObject", "@id": "agent_pypi_history.csv", |
| "name": "agent_pypi_history.csv", "contentUrl": "data/csv/agent_pypi_history.csv", |
| "encodingFormat": "text/csv", |
| "description": "PyPI/npm download history per agent (1,440 rows)."}, |
| {"@type": "cr:FileObject", "@id": "devtools_signals.csv", |
| "name": "devtools_signals.csv", "contentUrl": "data/csv/devtools_signals.csv", |
| "encodingFormat": "text/csv", |
| "description": "VS Code Marketplace install counts (89 rows)."}, |
| {"@type": "cr:FileObject", "@id": "agent_registry.csv", |
| "name": "agent_registry.csv", "contentUrl": "data/csv/agent_registry.csv", |
| "encodingFormat": "text/csv", |
| "description": "The 50-agent registry: name, category, GitHub repo, package names, marketplace IDs, search terms (50 rows)."} |
| ] |
| } |
|
|
