AI-Research Agents in the Wild. From GitHub and arXiv to Regularities and Gaps

1 Introduction

• The public record mixes runnable agents, benchmarks, libraries, evaluation harnesses, and other artifacts, so repository counts do not directly measure runnable-agent prevalence.The paper treats role coding and an explicit sampling frame as necessary for estimating that prevalence.Source: §1 · p. 2
• The study asks how nine promoted lineages relate to the 139-record registry, how six pre-specified regularities fare on 25 new repositories, and how three relations connect papers and repositories.The third question concerns curator-assigned relatedness, companion code, and confirmed authorship relations.Source: §1 · p. 2
• R1–R4 receive bounded support, the broad form of R5 is contradicted, and R6 remains exploratory after its directional prediction fails.R1–R4 concern design constraints, bounded programmatic scoring, an empty design cell, and failure-pattern rankings; R5 concerns cross-lineage patterns and R6 pattern growth.Source: §1 · p. 2
• Its design theory states six candidate regularities with explicit trigger observations and subjects them to a prospective case-level audit.The audit records where the theory survived, where its scope changed, and which questions remain open.Source: §1 · p. 2
• The evidence-grounded map separates repository-registry size, lineage membership, artifact role, and recurring antipatterns.These distinctions keep public-artifact counts separate from lineage and role classifications.Source: §1 · p. 2
• The cross-corpus analysis separately measures curator-assigned relatedness, companion code, and authorship overlap, including a corrected lower bound on participation across software and papers.These separate denominators are intended to make code-availability and community claims explicit for reviewers.Source: §1 · p. 2

2 Corpus and Method

• Both registries use one agent-drafted evidence card per record, with human authors setting scope, adjudicating difficult cases, and reviewing the source-anchored cards.Aggregate counts are regenerated from structured fields rather than card prose.Source: §2 · p. 4

2.1 Repository registry

• The repository registry freezes 139 canonical records, but it is an engineered discovery sample rather than a prevalence sample of runnable agents.The records span agents, benchmarks, libraries, substrates, meta-resources, companions, and boundary artifacts; estimating runnable-agent prevalence requires role coding and an explicit sampling frame.Source: §2.1 · p. 4

2.2 Paper registry and relation types

• The paper registry contains 101 records for 98 unique bibliographic artifacts, and inaccessible full text for 25 records makes observed paper-side link coverage a lower bound.Corpus-relatedness is curator-assigned, companion-code records associated public code, and authorship relations distinguish byline matches from commit-adjudicated identity matches; a parser must accept both stored list formats to reproduce the headline relatedness counts.Source: §2.2 · pp. 4–5

2.3 Lineages and coding primitives

• Nine promoted lineages yield 59 memberships among 52 repositories, with seven repositories assigned to two lineages; coding also distinguishes five selection signals and four judge types.The coding records a hybrid judge separately; 19 of 139 records use it, while 38 records have neither a candidate comparison nor an acceptance gate. Promotion counts include single-artifact records, so pattern totals are not counts of twice-confirmed regularities.Source: §2.3 · pp. 5–6

2.4 Time-stamped prospective protocol

• Six regularities and minimum trigger observations were fixed before auditing 25 repositories collected in June 2026, with trigger outcomes separated from corpus-wide counterevidence and post-hoc restrictions.The baseline registry held 76 canonical records; growth to 139 also includes material outside the prospective batch. Table 1 reports audited-batch observations, and P4 uses an imported ranking based on a four-lineage subset.Source: §2.4 · p. 6, Table 1 · p. 7

2.5 Artifact-level verification of the codings

• Artifact checks resolved 3,306 of 3,547 pinned references to verbatim source lines, while only 230 anchors were cited inline beside the claims they support.Across 176 bindings for selection signal, judge type, loop topology, and tool category, no contradictions were found; pattern membership and iteration-cost coordinates had 10 contradicted bindings each among 64 and 59, respectively.Source: §2.5 · pp. 7–8

2.6 Authorship audit and reproducibility

• Identity matching starts from normalized display-name candidates but requires explicit evidence; commit-adjudicated matches use complete commit histories and alone support lower-bound claims.The paper reports results at identity, author-slot, and paper levels, keeping the byline-matched tier separate because it bounds repository authorship in neither direction.Source: §2.6 · p. 8

3.1 Six candidate regularities

• The tested seed-line case used programmatic acceptance, while no audited case combined sub-10-minute iteration with an open-ended artifact under manual coding.These observations left the specified R1 and R2 triggers unobserved in the 25-case batch.Source: §3.1 · p. 9, §3.2 · p. 10
• Across 139 repositories, R2 is rule-dependent: observed durations leave the forbidden cell empty without measurements in its 17 open-ended repositories, while configured bounds yield two cases.The paper reports both readings rather than selecting the one favorable to the hypothesis.Source: §3.1 · p. 9
• Broader evidence contradicts R5’s original cross-lineage scope: uncalibrated judges occur in four lineages, while architecture patterns appear in at most two.The architecture-only restriction followed observation of the counterexample and is post hoc.Source: §3.1 · p. 10
• R6’s directional prediction failed: the runtime-contract cluster stayed at three instances across 15 observations, leaving the growth regularity exploratory.The audit also found no long-stagnant two-instance pattern jumping to five.Source: §3.1 · p. 10, §3.2 · p. 11
• R4’s failure-pattern lead is bounded to this corpus and coding: under nine promoted lineages, an architecture pattern leads in two of four weightings.The prospective audit’s imported four-lineage ranking retained a failure pattern in first place, but ranks two through five were not recomputed.Source: §3.1 · p. 9, §3.2 · p. 11

3.2 Prospective outcomes

• Zero of six specified trigger events appeared in 25 audited repositories; R1–R4 receive bounded support, R5 required a post-hoc restriction, and R6 remains exploratory.The authors caution that calling all six regularities confirmed would exceed the evidence.Source: §3.2 · pp. 10–11

3.3 What the theory currently explains

• The theory currently treats structural commitments, bounded measurable artifacts, cross-lineage evaluation failures, and tournament aggregation as useful mechanisms, but its ranking measures corpus support—not system performance.Independent data are needed to test generality; future cases can populate a cell, reverse a rank, cross a lineage boundary, or change a growth trajectory.Source: §3.3 · p. 11

3.4 Post-freeze re-test of the empty design cell

• Post-freeze re-testing left R3 standing on whether the Pareto front must feed a later loop, while broadening its explanation to three routes to the empty cell.The routes are evaluation-cost refusal, tournament aggregation, and shipped-but-unreachable selection code.Source: §3.4 · p. 12

4.1 Repository ecology and lineage coverage

• Nine promoted lineages cover 52 of 139 repositories (37.4%), with 59 memberships because seven repositories belong to two lineages; 87 records have no promoted-lineage membership.The lineage map is an overlapping subset of the registry, and role-specific claims require role-specific denominators.Source: §4.1 · p. 12

4.2 Curator-assigned relatedness graph

• Sixty-four of 101 papers have a curator-assigned relatedness edge, reaching 23 of 139 repositories; the observed paper-side rate is a lower bound of 63.4%–88.1% under adversarial imputation.The 212 edges are concentrated: six repositories contribute 120 (56.6%), with the concentration attributed primarily to curation choices and familiarity.Source: §4.2 · pp. 12–13

4.3 Companion code and authorship overlap

• Fifteen papers link to companion code resolving to a canonical repository, and at least ten of those papers have an author confirmed on that repository.The evidence ledger confirms at least 58 paper authors who also author canonical registry repositories, across 63 author slots in 15 papers; 50 linked papers have no confirmed registry-author identity.Source: §4.3 · p. 13, Table 3 · p. 13

4.4 Interpretation

• The graph measures curator-assigned relatedness, not citation behavior, and the confirmed author overlap remains a lower bound rather than a community-structure estimate.Private use lies outside both sampling frames and remains unmeasured.Source: §4.4 · p. 14

5.1 Scientific value and novelty

• The paper tests explicit design triggers and keeps distinct outcomes visible: R1–R4 remain viable within a finite audit, R5 loses broad scope, and R6 loses its directional companion prediction.Its coding check also distinguishes source-demonstrable aggregates from coding families with contradictions.Source: §5.1 · p. 14

5.2 Implications for system design

• The design implications pair acceptance mechanisms with units of work, propose controlled comparisons of candidate-comparison schemes, and call for frozen snapshots to study pattern growth.The paper also recommends calibration reporting for language-model judges and treats cross-lineage failure transfer and evaluation safeguards as design concerns.Source: §5.2 · pp. 14–15

5.3 Implications for reviews of the field

• Reviews should declare their unit of analysis and evidence class: curator-assigned relatedness is not intellectual visibility, and identity counts depend on the identity model and evidence policy.The supported authorship result is a doubly adjudicated lower bound of 18 identities; the byline-matched tier of 58 is reported separately.Source: §5.3 · p. 15

5.4 Open problems

• Open questions include how to classify a near-match to the vector-signal/language-model-judge cell, why judge failures recur, and whether runtime-contract patterns grow beyond three observations.Community-structure estimates still require mention extraction, complete entity resolution, and role-stratified denominators; symmetric debate also lacks a completed prospective test.Source: §5.4 · p. 15

5.5 Relation to prior work

• Unlike capability-centered surveys and architecture accounts, this study analyzes public artifacts and recorded relations, combining repository evidence, canonical identities, curated paper links, and a temporal test.This provides an evidence-grounded complement to prior synthesis work.Source: §5.5 · p. 16

5.6 Limitations

• The repository registry is an engineered discovery sample, so its 139 records do not estimate global prevalence; artifact roles and several design variables also lack corpus-wide independent double coding.The lineage relation conflates ancestry, shared substrate, influence, and structural similarity, while the coding check covered only a subset and did not test second-coder agreement.Source: §5.6 · p. 16
• The curated paper registry includes 25 records without accessible full text, and a seven-paper check found locatable mentions for only three of 29 credited edges.The prospective audit used 25 non-random cases without external preregistration; authorship remains a lower bound, and repositories can change after inspection.Source: §5.6 · p. 16

6 Conclusion

• Nine promoted lineages contain 59 memberships among 52 repositories, overlapping on seven, within a registry of 139 canonical public records.The map distinguishes the broad artifact registry from the smaller design-lineage relation.Source: §6 · pp. 16–17
• The six candidate regularities retain mixed outcomes: R1–R4 remain supported within bounds, R5’s broad cross-lineage claim is contradicted, and R6 remains exploratory after a failed directional prediction.The outcomes define subsequent prospective tests.Source: §6 · p. 17
• The graph has 212 curator-assigned edges, but a full-text sample located mentions for only three of 29 checked edges; confirmed author overlap is reported alongside unresolved community structure.The contribution is a theory with explicit triggers and denominators that keeps unresolved cases visible rather than treating missing observations as positive evidence.Source: §6 · p. 17

Use of Generative AI

• Coding agents assisted with candidate triage, source navigation, analysis-code development, consistency checks, prose revision, and figure production, and drafted both registries.Paper records came from scripted language-model passes; repository records came from coding subagents reading pinned revisions.Source: Use of Generative AI · p. 17
• The authors report 63 of 144 compendium commits with model co-authorship trailers and two commits by copilot-swe-agent[bot].These figures are as of 26 July 2026.Source: Use of Generative AI · p. 17
• The frozen registries were stress-tested with a prospective audit of 25 new repositories and a separate post-freeze re-test of the R3 empty design cell.The re-test triaged 319 of 962 post-freeze repositories; eleven of fifteen independent re-checks refuted the trigger claims they were given, calibrating single-agent trigger reports.Source: Use of Generative AI · p. 17
• Headline counts were recomputed from agent-drafted fields, and pinned source anchors do not make those codings independent of their authors.No coding was double-coded by a second annotator when written; authors report re-deriving codings against artifacts and disclose contradictions.Source: Use of Generative AI · pp. 17–18
• Human authors selected the research questions, adjudicated scientific interpretations, reviewed the manuscript, and retain responsibility for its content.The evidence cards preserve resolvable anchors, but that auditability does not make the original coding independent.Source: Use of Generative AI · p. 18

Data and Code Availability

• The versioned archival snapshot of the companion compendium is the identifier of record, not the working repository mirror or its date-dependent public reachability.It contains the registries, evidence cards, identity-adjudication ledger, analysis programs, generated tables, figure source, relation tables, and manuscript source.Source: Data and Code Availability · p. 18
• Supplementary Tables S1–S3 reproduce the full repository and paper registries and the named-tool, companion-code, and confirmed-author relations.The article gives the composite pattern-strength weighting, bootstrap specification, and iteration-cost coding rule and matrix; the deposit adds per-repository evidence quotes.Source: Data and Code Availability · p. 18

A Metric Definitions and Provenance

• Table 4 presents denominators and evidence objects for the headline results.Source: Table 4 · p. 19

A.1 Pattern-strength weighting sensitivity

• The sensitivity ranking uses a 76-record baseline, and ranks two through five have not been recomputed against the enlarged corpus.A spot check raised judge-uncalibrated from 16.5 to about 17.95 on at least 11 repositories instead of 8, and keep-or-revert from 12.8 to about 13.22 on 14 instead of 13.Source: §A.1 · p. 19
• The imported score counts lineage support across four lineages, whereas the paper’s other lineage results count memberships across nine promoted lineages.For keep-or-revert, the imported relation records one lineage, while the nine-lineage relation places its supporting repositories in three.Source: §A.1 · p. 19
• Four defensible weightings combine tools, papers, and lineage inputs, with Tier S defined by the top-cluster gap in each variant.The inputs are printed with the scores for recomputation.Source: §A.1 · pp. 19–20
• The leading pattern changes with the weighting and lineage relation: an architecture pattern leads under two nine-lineage weightings, while an antipattern leads under the other two and all four imported-relation weightings.Under the nine promoted-lineage relation, the architecture pattern scores 16.8 against 16.5 with baseline weighting and 19.0 against 16.0 with linear tools; the antipattern scores 11.7 against 10.2 with unweighted lineages and 20.5 against 16.8 with paper-heavy weighting.Source: Table 5 · p. 20, §A.1 · p. 20
• The top-five pattern identities persist across all four weightings and both lineage relations, while their within-cluster ordering and leading position do not.The top cluster contains two antipatterns, two architecture patterns, and one component under every variant; its ranking is not an objective to optimize.Source: Table 5 · p. 20, §A.1 · p. 20
• Correcting keep-or-revert’s support changes its baseline score to 14.8 and restores the antipattern’s lead, while the architecture pattern leads under linear tools at 17.0.The correction removes ML-Agent, whose update-in-place gradient descent keeps no candidate pool to keep or revert, and adds NanoResearch; the corrected support spans two lineages.Source: §A.1 · p. 20

A.2 Relatedness-graph uncertainty estimates

• B=10,000 bootstrap replicates with seed 20260725 resample 101 paper records for paper-side rates and 139 canonical repositories for repository-side rates.The procedure is specified so the intervals can be regenerated from Supplementary Tables S1–S3 without deposit code.Source: §A.2 · pp. 20–21
• Top-6 stability counts how many of the observed six most-linked repositories reappear in each paper-level resample.The six are Agent Laboratory, The AI Scientist, MLAgentBench, AI-Researcher, NanoResearch, and ResearchAgent.Source: §A.2 · pp. 20–21
• For 25 papers without retrievable full text, adversarial imputation bounds the paper-side rate from 63.4% to 88.1%.The floor assumes zero edges for those records; the ceiling assumes at least one edge per record.Source: §A.2 · p. 21

A.3 Cost-versus-ambition coverage matrix

• Wall-clock entries distinguish observed durations from configured ceilings; per-step, per-query, or per-experiment figures bound an iteration only when that unit contains one.No numeric imputation is performed.Source: §A.3 · p. 21
• Re-coding 139 rows yields 14 measured durations, 56 configured ceilings, and 17 open-ended repositories, versus 6 measurements and 7 open-ended repositories previously reported.Four of the six earlier measurement anchors were ceilings or values absent from the cited source.Source: §A.3 · p. 21
• 0 of 17 open-ended rows have measurements under rule (a), while two repositories occupy the under-600-second open-ended cell under rule (b).Under rule (a), the empty cell reflects no measurements rather than a measured absence; under rule (b), two bounds apply to whole executions or runs.Source: Table 7 · p. 22, §A.3 · p. 22
• R2 remains conditional: observed-duration-only evidence is uninformative, whereas counting configured bounds yields counterexamples; published per-iteration durations could settle the question.No repository currently publishes measured per-iteration durations in the open-ended column.Source: §A.3 · p. 22

B Authorship Evidence Policy

• Authorship cases require explicit evidence such as a companion repository, contributor record, adjudicated identity, or documented cross-paper mapping; matching the paper's own byline transcription is excluded from lower bounds.Four registry records of the excluded kind account for 40 of 63 byline-matched slots.Source: B Authorship Evidence Policy · pp. 22–23
• Accepted cases are checked against complete commit histories across all refs; only a corroborated author or committer identity qualifies as commit-adjudicated.Handle resemblance without corroborating email or profile evidence is reported separately; histories also check whether a repository can support the byline-matched slots charged to it.Source: B Authorship Evidence Policy · p. 23
• Conflicting affiliations, common names, and unresolved handles remain open, and byline-matched counts bound repository authorship in neither direction.Unresolved cases remain unclassified with respect to community alignment.Source: B Authorship Evidence Policy · p. 23

From GitHub and arXiv to Regularities and Gaps

• The supplement lists 10 June 2026 as the analytic freeze and names Aleksey Komissarov and Andrey Ustyuzhanin.Source: From GitHub and arXiv to Regularities… · p. 25

Scope and relation definitions

• The registries enumerate 139 repositories and 101 paper records resolving to 98 bibliographic artifacts; three paired records retain separate identifiers.Repository IDs are assigned lexicographically within the freeze, while paper IDs preserve curated record identifiers.Source: Scope and relation definitions · p. 25
• Named-tool, companion-code, and confirmed-author relations respectively capture a paper naming or discussing a repository, its code URL resolving to one, and explicit evidence of shared authorship.Blank cells mean a relation is unobserved or unresolved under the protocol, not that the populations are separate.Source: Scope and relation definitions · p. 25

S1 Complete canonical repository registry

• 139 canonical public repository records appear in the complete registry at the analytic freeze.The table’s final three columns list paper identifiers for named-tool, companion-code, and confirmed-author relations, respectively.Source: Table S1 · p. 25
• The segment lists OpenRaiser/NanoResearch as a multi-agent-scientist record and snap-stanford/MLAgentBench [1] as an mlagentbench-derived record, each followed by associated P-coded identifiers.Source: S1 Complete canonical repository… · pp. 28–29
• The cited segment assigns G097 to multi-agent-scientist, G112 to clune-lab-open-endedness, and G104 and G106 to seed-line, documenting three lineage labels across four listed entries.Source: S1 Complete canonical repository… · pp. 28–29
• Some repository records also carry associated paper IDs, including P017 for OPPO-PersonalAI/FINDER_DEFT and P044 for RUC-NLPIR/WebThinker.The segment also lists multiple paper IDs alongside repositories such as AI-Scientist and MLAgentBench.Source: S1 Complete canonical repository… · pp. 28–29

• Table S2 contains all 101 paper records at the analytic freeze; repository IDs resolve through Table S1, and author entries count evidence-confirmed author slots for each paper–repository pair.Source: Table S2 · p. 30
• The cited records include P013, P014, P017, and P064; P013 is marked code cited, P017 code shipped, P014 full text, and P064 paper only.Source: S2 Complete paper registry · pp. 31, 33

S2 Complete paper registry

• Several records carry curator-assigned links to repository IDs, including G051, G064, G111, and G032.The listed related-repository sets vary by paper; for example, P094 is linked to G111, while P100 is linked to G053, G097, G102, and G011.Source: S2 Complete paper registry · pp. 33–35
• The cited entries show code-shipped, code-cited, demo-only, full-text, and unavailable records: P067 is code shipped, P080 code cited, P086 demo only, P075 full text, and another record unavailable.Source: S2 Complete paper registry · pp. 33–34
• The segment flags unresolved identity for P068 and P069, and lists the same drone-inspection paper under P081 and P088.The repeated P081/P088 entry gives the same title and identifier, 2025; 2510.00259.Source: S2 Complete paper registry · p. 33

S3 All observed paper–repository relations

• Table S3 defines three relation types—named-tool (N), canonical companion-code (C), and evidence-confirmed paper-author/repository-author (A)—and allows multiple types for one pair.Multiple letters on a row denote distinct evidence for the same paper–repository pair.Source: Table S3 · p. 35
• P013 Deep Research: A Systematic Survey is linked to mangopy/deep-research-survey by a canonical companion-code edge (C).Source: S3 All observed paper–repository… · p. 36
• P015 DR.BENCH is linked to evigbyen/drbench by a canonical companion-code edge (C).Source: S3 All observed paper–repository… · p. 37
• P018 DualResearch is linked to Alpha-Innovator/InternAgent through an evidence-confirmed author relation (A) naming Bo Zhang.Source: S3 All observed paper–repository… · p. 37
• COMPASS (P025) is paired with HKUDS/AI-Researcher, Human-Agent-Society/CORAL, and SakanaAI/AI-Scientist-v2.Source: S3 All observed paper–repository… · p. 38
• Marco DeepResearch (P028) is paired with AIDC-AI/Marco-DeepResearch, HKUDS/AI-Researcher, and SakanaAI/AI-Scientist-v2.Source: S3 All observed paper–repository… · p. 38
• What Makes AI Research Replicable? (P034) is paired with SakanaAI/AI-Scientist, du-nlp-lab/MLR-Copilot, and snap-stanford/MLAgentBench.Source: S3 All observed paper–repository… · p. 39
• Can We Predict Before Executing Machine Learning Agents? (P038) is paired with MASWorks/ML-Agent, SakanaAI/AI-Scientist, and SamuelSchmidgall/AgentLaboratory.Source: S3 All observed paper–repository… · p. 39
• SAGE (P039) is paired with HKUDS/AI-Researcher, OpenRaiser/NanoResearch, and SakanaAI/AI-Scientist.Source: S3 All observed paper–repository… · p. 39
• P047, “UFO3: Weaving the Digital Agent Galaxy,” is paired with six repositories, from HKUDS/ClawTeam to vukrosic/auto-research.Source: S3 All observed paper–repository… · p. 40
• P048, “K-Dense Analyst,” is linked to HKUDS/AI-Researcher, OpenRaiser/NanoResearch, SakanaAI/AI-Scientist, AgentLaboratory, and AutoResearchClaw.Source: S3 All observed paper–repository… · pp. 40–41
• P049, “Rethinking Memory Mechanisms of Foundation Agents in the Second Half,” is paired with repositories including ResearchAgent, AI-Scientist, AgentLaboratory, aideml, and awesome-agent-memory.Source: S3 All observed paper–repository… · p. 41
• P053, “Analytical Search,” is paired with HKUDS/AI-Researcher, OpenRaiser/NanoResearch, and SamuelSchmidgall/AgentLaboratory among its listed relations.Source: S3 All observed paper–repository… · p. 41
• P080, “Improving and Evaluating Open Deep Research Agents,” is paired with JinheonBaek/ResearchAgent, SamuelSchmidgall/AgentLaboratory, and nickscamara/open-deep-research.Source: S3 All observed paper–repository… · p. 44
• P081, “A Hierarchical Agentic Framework for Autonomous…,” is paired with HKUDS/ClawTeam, Human-Agent-Society/CORAL, and SamuelSchmidgall/AgentLaboratory.Source: S3 All observed paper–repository… · p. 44
• P083, “MedCoAct: Confidence-Aware Multi-Agent…,” is paired with Human-Agent-Society/CORAL, JinheonBaek/ResearchAgent, and PouriaRouzrokh/LatteReview.Source: S3 All observed paper–repository… · p. 44
• P084, “Autonomous Data Processing using Meta-Agents,” is paired with RightNow-AI/autokernel, WecoAI/aideml, du-nlp-lab/MLR-Copilot, and snap-stanford/MLAgentBench.Source: S3 All observed paper–repository… · p. 44
• The relation listing includes “Dual AI-Agent Collaboration” alongside SamuelSchmidgall/AgentLaboratory.Source: S3 All observed paper–repository… · p. 44

Corpus sources

• Entries include karpathy/autoresearch, HKUDS/AutoAgent, and InternScience/MLEvolve.These are examples from the repository records listed in this part.Source: Corpus sources · p. 48
• The list also includes openags/Auto-Research, microsoft/RD-Agent, and SakanaAI/AI-Scientist-v2.Each is listed as a GitHub repository accessed 10 June 2026.Source: Corpus sources · pp. 49–50
• Other listed records include RUC-NLPIR/Search-o1, Tencent/CognitiveKernel-Pro, and trevin-creator/autoresearch-mlx.These entries likewise identify GitHub repositories accessed 10 June 2026.Source: Corpus sources · pp. 50–51
• Companion artifacts are listed for Memento [P001], the agentic multimodal-LLM survey [P008], and Deep Research: A Systematic Survey [P013].The records name the Memento, Awesome-Agentic-MLLMs, and Deep-Research-Survey repositories, respectively.Source: Corpus sources · pp. 51–52
• Code or companion artifacts are listed for DR.BENCH [P015], the deep research agents study [P017], and interactive theorem proving research [P027].The records point to drbench, FINDER_DEFT, and MathResearchPrompts, respectively.Source: Corpus sources · p. 52
• Companion artifacts are also listed for Marco DeepResearch [P028] and the paper on executable knowledge graphs [P034].The entries name Marco-DeepResearch and zjunlp/xKG, respectively.Source: Corpus sources · p. 53
• Bibliographic metadata remains unresolved for [P055] and [P069].Source: Corpus sources · pp. 54–55
• [P060] VSearcher lists a GitHub companion artifact, while [P049] links to an AgentMemoryWorld repository.Source: Corpus sources · p. 54