Autonomous Agents Coordinating Distributed Discovery Through Emergent Artifact Exchange

• Most AI systems remain interactive, responding to human prompts, while scientific discovery involves iterative exploration, tool use, hypothesis testing, and comparison across evidence.The introduction motivates autonomous investigation that can solicit input across diverse capabilities and converge across independent lines of reasoning.Source: §1 · pp. 1–2
• The framework emphasizes a persistent multi-agent ecosystem rather than human assistance or automation of a single research pipeline.Agents can chain tools, publish structured artifacts, and engage with one another’s findings as part of distributed discovery.Source: §1 · p. 2
• SCIENCECLAW + INFINITE supports distributed autonomous exploration through tool-chaining agents that produce versioned artifacts linked to the findings they support.SCIENCECLAW provides the computational layer and interoperable scientific tools; the artifacts preserve explicit provenance chains from raw outputs to findings.Source: §1 · p. 2
• INFINITE provides structured scientific discourse where agents share findings, supporting artifacts, and open questions, enabling posts to prompt follow-up analyses.Humans and other agents can evaluate posts, creating a feedback loop between published results and subsequent investigation.Source: §1 · p. 2
• Four investigations span SSTR2 peptide design, impact-resistant ceramic screening, resonance-inspired cross-domain design, and quantifiable analogy between disconnected domains.The SSTR2 study demonstrates convergence among independent agents; resonance-inspired candidates are validated through physics-based analysis, while the final investigation yields an explicit, quantifiable representation for evaluation.Source: §1 · p. 2

2.1 System Overview

• Investigations culminate in structured posts and, when provenance and quantitative results suffice, self-contained reports with figures; community feedback helps shape subsequent research directions.Published investigations become visible to peer agents and the broader community, whose votes and actions feed new need signals into the cycle.Source: §2.1 · p. 3

2.2.3 Investigation Pipeline

• Distinct scientific profiles shape agents’ reasoning, while an open skill registry and JSON-based tool chains let them select and sequence investigations without hardcoded routing.A synthesis pass uses the sequential skill outputs to form a testable hypothesis, identify cross-database convergences, and draft a finding narrative.Source: §2.2.1 · p. 4, §2.2.2 · p. 4, §2.2.3 · p. 4

2.2.4 Artifact Layer and Provenance

• Each skill invocation creates an immutable, typed artifact with a content hash and parent lineage, forming a DAG that lets published claims be traced to their underlying computations.Artifacts also carry quality flags and needs; the DAG supports integrity checks and reproducible handoffs across tools.Source: §2.2.4 · pp. 4–5

2.2.5 ArtifactReactor: Emergent Cross-Agent Coordination

• The ArtifactReactor coordinates agents without central task assignment by matching broadcast needs and compatible payloads, prioritizing work, and synthesizing multi-parent artifacts.Pressure scoring weights novelty, centrality, depth, and age; mutation mechanisms address redundancy, stagnation, and conflicting findings.Source: §2.1 · p. 3, §2.2.5 · pp. 5, 7

2.2.6 Persistent Memory

• Persistent journals, investigation trackers, and knowledge graphs let agents build on prior cycles, while a six-hour heartbeat runs investigations and records auditable activity.Human redirects take precedence over normal scoring, and compatible peers can join advertised sessions, claim subtasks, and contribute artifacts to a joint finding.Source: §2.2.6 · p. 7, §2.2.7 · pp. 7–8

2.3.1 Data Model

• INFINITE presents findings as structured scientific posts linked to artifact metadata, with typed relations and threaded actions that make evidence and discourse machine-readable.Post fields include hypotheses, methods, findings, data sources, and open questions; peer links can cite, contradict, extend, or replicate findings.Source: §2.3 · p. 8, §2.3.1 · p. 8

2.3.2 Governance and Incentive Design

• INFINITE ties visible evidence and provenance to community engagement and reputation, while capability checks and rate limits constrain agent behavior.Registration requires proof of capability, and the passage specifies uniform limits of one action per 30 minutes, one per 20 seconds, and 50 per day.Source: §2.3.2 · p. 9

2.3.3 Agent–Platform Integration

• Agents connect to INFINITE through a typed API, poll community activity, and use chats or redirects to update their journals and prioritize human-specified questions.Redirected sub-questions bypass normal gap-detection scoring and move to the top of the next-cycle queue.Source: §2.3.3 · pp. 9–10

3.1 Quantitative Characterization of Autonomous Investigations

• Artifact generation ranged from 52 to 177 per investigation, synthesis density from 12% to 48%, and average DAG depth from 2.00 to 2.25.The studies ran autonomously across 8–13 agents; design-driven investigations had deeper dependency chains, while Resonance and Urban-Material used more sequential pipelines.Source: §3.1 · pp. 10–11

3.2 Protein Design

• Structural, evolutionary, and ESM-2 analyses converged on the K-T-C-centered motif as an SSTR2 binding anchor, with surrounding residues offering greater mutation flexibility.The mutation scan predicted improved sequence fitness for A1M and W8L substitutions, while mutations in the conserved CWKTCT-like motif were less tolerated.Source: §3.2 · pp. 11–12, Figure 7 · p. 12
• The optimized peptide MGLKNFFLKTFTSC improved predicted sequence fitness but weighed approximately 1639 Da, exceeding approved SSTR2-targeting therapeutics cited in the study.The authors suggest cyclization and sequence truncation as more promising directions; without docking or molecular dynamics, the findings remain hypothesis-generating rather than definitive binding-affinity predictions.Source: §3.2 · pp. 12–13
• The SSTR2 investigation exposed collaboration provenance through agent-interaction and tool-level dataflow views that trace exchanged artifacts from computational outputs to published conclusions.Approximately ten agents contributed analyses spanning structure retrieval, sequence alignment, ESM scoring, physicochemical profiling, and candidate ranking.Source: §3.2 · p. 13

3.3 Materials Discovery

• The ceramic workflow combined database screening, stability analysis, and synthesis-feasibility assessment, identifying less-explored stable candidates Mg2B24C and MgB9N with estimated synthesis success probabilities of 0.38 and 0.31.Neither has substantial experimental synthesis literature, and the authors note that predicted feasibility may reflect both practical synthesis challenges and limited documentation.Source: §3.3 · p. 15
• Of 212 ceramic phases screened, 14 met the density and bulk-modulus constraints, and 7 were stable on the convex hull, including B4C and B6O.B4C had density 2.54 g/cm3 and bulk modulus 238 GPa; B6O had density 2.62 g/cm3 and bulk modulus 229 GPa. Both lay above the dataset’s density–stiffness trend.Source: §3.3 · p. 15, Figure 9 · p. 15

3.4 Cross-Domain Resonance

• Across 39 resonators, PCA explained 88.6% of variance in two components, and agents identified a design-space gap about 12 times closer to biology than engineered materials.Bach chorale motif characteristics mapped onto acoustic periodicity, while the gap reflected missing materials combining high membrane character and hierarchy; ANOVA supported domain differences.Source: §3.4 · p. 16, Figure 10 · p. 17
• Agents specified a Hierarchical Ribbed Membrane Lattice to target the resonance gap, then used geometric realization and 3D finite-element modal analysis for physical validation.The proposed design targeted 2–8 kHz, with predicted PCA coordinates (−0.50, +0.45) and membrane_character ≈0.85 and hierarchical_layering ≈0.90.Source: §3.4 · p. 16
• 2.116 kHz fundamental frequency, nine in-band elastic modes, and 1.5 modes/kHz modal density made v1_cricket_fine the best candidate, matching cricket-wing measurements (Prestwich & Walker 1981: ≈1–2 modes/kHz).The candidate uses a cricket-harp topology with 1.667 mm fine secondary rib spacing.Source: §3.4 · p. 17
• The full 3D model predicted f1 = 2116 Hz, a 148 Hz (7%) upward shift from the 2D model’s 1968 Hz, bringing the fundamental into band without clamped boundaries.The shift was attributed to volumetric rib-frame stiffening absent from the 2D continuum approximation.Source: §3.4 · p. 17
• Gap detection placed the design gap at (−0.521, +0.425), 12× closer to biology than current materials, motivating a Hierarchical Ribbed Membrane Lattice candidate.Agents mapped music, biology, and materials into a shared resonance feature space, with musical interval patterns informing phononic-crystal periodicity.Source: §3.4 · p. 18

3.5 Formal Analogy Between Urban Morphology and Grain Boundary Evolution

• A review recovered 18 papers in non-citing urban-physics and materials communities; agents built a 66-concept ontology with 9 cross-domain edges, leaving assemblage and cyborg-urban unmapped.Mapped pairs included block–grain, street–boundary, and junction–triple junction.Source: §3.5 · pp. 18–19, Figure 12 · p. 19
• On a normalized 60-point dataset, power-law fitting gave α = 1.25±0.08 with R2 = 0.71, while OLS gave β = 0.83, p = 0.002, and R2 = 0.68.The literature-derived growth metrics support compatible scaling descriptions, not a shared mechanistic law.Source: §3.5 · p. 19
• The reduced urban and grain graphs shared degree sequence [3, 3, 3, 2, 2, 1, 1] and were not separated by tests, but seven-node networks limit statistical power.They also had nearly identical transitivity and the same Weisfeiler–Leman hash; Bayesian comparison assigned 0.82 probability that their exponents differ by less than 0.5.Source: §3.5 · p. 19, Figure 13 · p. 20
• The agents compressed the cross-domain analogy into an executable six-rule L-system grammar that grew from 3 terminal segments at step 1 to 805 at step 5.Its symbols encode settlement or nucleation, growth fronts, boundaries, infill, junctions, and consolidation or coarsening.Source: §3.5 · pp. 19–20, Figure 14 · p. 20
• The urban–grain result is a constrained formal analogy, not mathematical isomorphism or shared governing equations; critics noted objective-function and timescale mismatches and hub-dependent graph brittleness.The hub mapping’s removal disconnects the network; the investigation formalizes an inspectable, executable, and falsifiable hypothesis rather than producing a material candidate.Source: §3.5 · p. 20

4 Conclusions

• SCIENCECLAW + INFINITE combines an extensible skill registry, lineage-preserving artifacts, and structured discourse to sustain decentralized scientific investigation across agents and cycles.Agents chain tools without hardcoded routing; need signals and schema overlap identify useful information, while autonomous mutation prunes redundant or stagnant workflows.Source: §4 · p. 21
• INFINITE makes computational findings inspectable through structured posts, visible provenance, and typed discourse relations that can guide later investigation cycles.Relations include cite, contradict, extend, and replicate; community actions and redirects support review, refinement, and follow-up.Source: §4 · p. 21
• Across four case studies, the framework enabled heterogeneous tool chaining, cross-domain synthesis, and outputs that remained traceable, testable, and open to falsification.The studies covered peptide analysis, materials screening, resonance across biology, materials, and physics, and a formal analogy between urban morphology and grain-boundary evolution.Source: §4 · p. 21
• Persistent artifact chains and community feedback enable cumulative participation, while new skills and participants can expand investigative pathways without redesigning the reasoning layer.Unresolved questions can become starting points for new work, and findings remain available for revisiting, reinterpretation, and extension.Source: §4 · p. 21

5 Materials and Methods

• The implementation combines the ScienceClaw agent framework, Infinite collaboration platform, and computational tools for independent investigations and sharing findings.Source: §5 · p. 21

Language Models

• Language models support topic analysis, skill selection, hypothesis generation, synthesis, and self-critique, with Claude Opus 4.6/4.5 as the primary backend.The LLM backend is configurable.Source: Language Models · p. 21
• The active LLM backend is set in llm_config.json and can be changed through the ANTHROPIC_MODEL, OPENAI_MODEL, or HF_MODEL environment variable.Source: Language Models · p. 22

ScienceClaw Agent System

• Agents use persistent agent_profile.json files for identity, research interests, and preferred tools, created through an interactive wizard or quick setup.The wizard collects a name, biography, expertise preset, research interests, preferred organisms for biology agents, and tools selected from over 200 skills.Source: ScienceClaw Agent System · p. 22
• Expertise presets cover biology, chemistry, materials science, and mixed profiles, defining initial interests, targets, and preferred tools.Examples include BLAST, PubMed, and UniProt for biology, and PubChem, ChEMBL, and TDC for chemistry.Source: ScienceClaw Agent System · p. 22
• Initialization generates the agent profile, SOUL.md personality file, and LLM configuration; provided Infinite credentials enable platform registration and an API key for authenticated posting.Source: ScienceClaw Agent System · p. 22

Memory and Persistence Layer

• Across cycles, agents retain locally stored memory through three complementary stores: an append-only journal, active-investigation state, and a semantic knowledge graph.Together, the stores support building knowledge across heartbeat cycles.Source: Memory and Persistence Layer · p. 22
• AgentJournal is an append-only JSONL record of timestamped observations, hypotheses, experiments, and conclusions, with methods capturing sources, reasoning, parameters and results, and supporting evidence.Each entry includes a timestamp, type, content, and optional metadata.Source: Memory and Persistence Layer · p. 22
• InvestigationTracker preserves a topic, hypotheses, experiments, and results across cycles, allowing agents to resume work and avoid duplication; its API supports creation, updates, and completion.Source: Memory and Persistence Layer · p. 22
• KnowledgeGraph represents scientific concepts as nodes and semantic relationships as edges, enabling agents to reason about connections and identify novel links across scientific areas.Nodes can represent proteins, compounds, diseases, processes, or findings; edges include activates, inhibits, associated-with, and contradicts.Source: Memory and Persistence Layer · p. 22

Autonomous Investigation Loop (6-Hour Cycles)

• The heartbeat daemon runs continuously and triggers investigation every six hours (21600 seconds by default), loading the agent profile and tracking state and execution logs locally.Its files are agent_profile.json, heartbeat_state.json, and heartbeat_daemon.log.Source: Autonomous Investigation Loop (6-Hour… · p. 23
• Each six-hour cycle reads community posts to identify gaps, generates and ranks hypotheses by novelty, feasibility, and profile fit, then investigates, posts findings when credentials exist, and engages.The loop uses rule-based pattern matching and LLM-powered synthesis to generate candidate hypotheses.Source: Autonomous Investigation Loop (6-Hour… · p. 23

Deep Investigation System

• The investigation workflow analyzes a topic, selects up to five relevant skills, executes them in sequence, synthesizes findings, logs the process, and returns structured outputs.The returned object includes selected skills, tool outputs, synthesized findings, and suggested experimental validation approaches.Source: Deep Investigation System · p. 23
• The investigation entry point relies on LLM-based analysis and skill selection rather than hardcoded tool chains; direct class-method calls bypass selection and yield shallow, generic outputs.Source: Deep Investigation System · p. 23

Skills and Tools Registry

• The registry exposes 200+ computational tools across scientific domains through standardized Python scripts that accept command-line arguments and return chainable JSON output.Tools span literature search, bioinformatics, chemistry, materials science, and machine learning.Source: Skills and Tools Registry · p. 23
• The skill registry records tool names, descriptions, parameters, output schemas, and categories so agents can discover tools and match them to investigation goals.Families include literature, proteins, compounds, genomics, and machine learning.Source: Skills and Tools Registry · p. 23
• The skill executor invokes a selected script with command-line parameters, parses its JSON result, and uses registry metadata to track the artifact types tools produce.The passage illustrates this with a TDC ADMET prediction command using a SMILES string and model name.Source: Skills and Tools Registry · p. 23

Infinite Platform

• Infinite uses Next.js 14, PostgreSQL, and Drizzle ORM, with database tables for agents, communities, posts, comments, votes, links, and notifications.Agent records include identity, capability verification, karma and reputation, moderation status, incident counts, engagement counts, and timestamps.Source: Infinite Platform · p. 24
• Infinite posts store structured scientific findings and open questions, while threaded comments and typed post links support discussion, citation, contradiction, extension, and replication.Posts also track engagement and moderation or spam flags; comments can carry scientific fields, and notifications track mentions, votes, replies, and citations.Source: Infinite Platform · p. 24
• Agent authentication verifies capability proof at registration, issues an API key, and then uses hashed-key checks and signed JWTs to authenticate subsequent requests.The JWT includes the agent ID and capability list.Source: Infinite Platform · p. 24

Multi-Agent Coordination and Collaborative Sessions

• The SessionManager coordinates investigations through shared sessions stored as JSON files in each agent’s sessions directory.Sessions organize participating agents and their domains, shared tasks and tool requirements, results, synthesis findings, and creation and completion timestamps.Source: Multi-Agent Coordination and… · p. 24
• Agents join sessions when their preferred tools match task requirements, claim tasks atomically, execute them, record artifact-backed results, and mark tasks complete.When synthesis is enabled, the manager integrates results from all agents into findings and shares them using post_finding().Source: Multi-Agent Coordination and… · p. 24
• Given a research topic, the AutonomousOrchestrator analyzes needed expertise and tools, creates 2–5 agent specifications, and registers temporary profiles with Infinite.It then creates a session, derives tasks from the topic analysis, triggers agents to work, synthesizes their artifact-producing results, and posts a comprehensive community result.Source: Multi-Agent Coordination and… · pp. 24–25

Artifact System for Computational Provenance

• Each skill invocation creates an addressable artifact that records the skill’s result for a specific agent and investigation, providing computational provenance.Artifacts are stored as JSON objects with a UUID, artifact type, producer and skill names, schema version, unchanged skill-output payload, investigation ID, UTC timestamp, and SHA256 content hash.Source: Artifact System for Computational… · p. 25
• Each agent’s artifacts are appended to a local JSONL store and referenced with an artifact://{agent_name}/{artifact_id} address.The append-only storage pattern matches the memory journal.Source: Artifact System for Computational… · p. 25
• Session managers gate posted artifacts by checking whether each artifact type is allowed by the agent’s preferred_tools profile.The SKILL_DOMAIN_MAP maps 200+ skill names to artifact types; agents without preferred_tools are unrestricted, and synthesis and peer_validation artifacts are permitted for all agents.Source: Artifact System for Computational… · p. 25

ArtifactReactor: Reactive Coordination Mechanism

• ArtifactReactor detects unmet demands and compatibility signals in the artifact store, allowing agents to respond autonomously without explicit orchestration.The reactor automatically triggers skill execution to fulfill detected needs.Source: ArtifactReactor: Reactive Coordination… · p. 25

Need Signals: Explicit Demand Broadcasting

• Synthesis artifacts can broadcast up to two NeedItems specifying a requested artifact type, a targeted entity or search term, a rationale, and optional fulfillment guidance.A NeedItem can request up to six competing variants and include preferred skills and parameter variants; signals enter a lightweight global index containing IDs, types, and need text rather than payloads.Source: Need Signals: Explicit Demand… · p. 25

Schema-Overlap Matching: Implicit Supply Discovery

• Schema-overlap matching discovers unclaimed peer artifacts when a skill’s normalized input parameter names intersect an artifact’s normalized top-level payload keys.For skills with a --describe-schema flag, a secondary JSON-schema path extracts structured input_json_fields, supporting richer matching for skills that accept arbitrarily named data via --input-json.Source: Schema-Overlap Matching: Implicit… · p. 25

Pressure-Based Scoring of Open Needs

• Open needs are handled before schema-overlap reactions and ranked by score = 2.0 · novelty + 1.0 · centrality + 0.5 · depth + 0.2 · age.Novelty is 1/(1 + coverage), centrality counts related same-type needs across the investigation, depth measures DAG edges to the root, and age is log(1 + age_in_minutes).Source: Pressure-Based Scoring of Open Needs · p. 26
• Unfulfilled needs receive greater novelty scores, while overlapping same-type requests across agents raise centrality and favor convergent demand over isolated requests.A need fulfilled zero times scores 1.0 for novelty; after two fulfillments it scores 0.33. Deeper derivations are preferred, and the logarithmic age term helps prevent starvation of old unanswered needs.Source: Pressure-Based Scoring of Open Needs · p. 26
• Agents cannot fulfill their own needs, and consumed needs are recorded by artifact, need, and variant IDs to prevent duplicate work.The record format is artifact_id:need_index:variant_id in consumed_needs.txt.Source: Pressure-Based Scoring of Open Needs · p. 26

Multi-Parent Synthesis and Cross-Agent Credit

• Compatible artifacts are merged oldest-to-newest into a shared-skill synthesis artifact with multiple parent IDs, recording every contributing agent in the DAG lineage.When parent keys overlap, the newest value overwrites; synthesis and peer_validation artifact types are always permitted across domain restrictions.Source: Multi-Parent Synthesis and Cross-Agent… · p. 26

Loop Prevention and Investigation Isolation

• Consumed artifact IDs prevent repeat reactions, self-cycles are blocked, and an optional investigation filter isolates scans when agent identities are reused.The filter scopes scans to one investigation ID, preventing cross-run pollution in a shared global index.Source: Loop Prevention and Investigation… · p. 26

Mutation Layer: Topology Self-Modification

• The ArtifactMutator forks stagnant artifacts, merges redundant siblings, and resolves conflicting siblings by cycle-safe grafting or synthesis.Stagnation means more than K cycles without children; redundancy means siblings share more than P% of keys; conflicts involve the same key with different values.Source: Mutation Layer: Topology… · p. 26
• Mutation policies set stagnation_cycles to 3, redundancy_threshold to 0.7, and max_mutations_per_cycle to 2 by default; thresholds drift with observed conflict and redundancy rates.These first-class mutation_policy artifacts let the agent population self-tune its exploration strategy without human intervention.Source: Mutation Layer: Topology… · p. 26

ArtifactReactor Implementation

• ArtifactReactor prioritizes need-driven reactions, then multi-parent synthesis and single-parent transforms, with optional mutation detection; its API also scans compatible artifacts and fulfillable needs.react(limit=3, investigation_id="") runs up to the specified limit; react_to_needs(limit=K) fulfills high-scoring open needs.Source: ArtifactReactor Implementation · p. 26
• Session-scoped caches load skill parameters and schemas once per session and consumed artifact and need IDs once per cycle, reducing filesystem I/O while preserving correctness for append-only JSONL stores.The reactor also handles compatibility checks, skill selection and execution, artifact creation, and consumption tracking.Source: ArtifactReactor Implementation · p. 26

Post Generation and Platform Integration

• Posts package each investigation as a title, synthesized narrative, hypothesis, methods, findings, sources, open questions, and tools used.Sources can include PMIDs, UniProt accessions, or database references.Source: Post Generation and Platform Integration · p. 27
• Posts must demonstrate multi-tool cross-validation, novel mechanistic insight, quantitative findings with confidence metrics or error bounds, and attribution to an authenticated agent with provenance tracking.Source: Post Generation and Platform Integration · p. 27
• The Infinite client validates authentication, rate limits, and karma thresholds before submitting posts; quotas and spam enforcement can restrict visibility or posting privileges.Limits include one post per 30 minutes, 50 comments per day, and 200 daily votes—or 400 for trusted agents; repeated low-quality or duplicate posts can trigger shadowbans or bans.Source: Post Generation and Platform Integration · p. 27
• Agents can add threaded comments with mentions, link posts as citing, contradicting, extending, or replicating one another, and retrieve notifications for mentions, replies, upvotes, and citations.Karma tiers determine privileges: trusted agents with karma of at least 200 and reputation of at least 1000 can moderate communities and influence governance.Source: Post Generation and Platform Integration · p. 27

Scientific Reasoning and Hypothesis Generation

• The scientific reasoning engine orchestrates five components, beginning with a GapDetector that ranks knowledge gaps by frequency, recency, and fit with agent interests.It draws on agent journals and community posts to identify incomplete understanding and questions raised by other agents.Source: Scientific Reasoning and Hypothesis… · p. 27
• The HypothesisGenerator uses prior investigations and LLM synthesis to turn gaps into mechanistic predictions with specified variables and predicted relationships.It formats these predictions as clear research questions rather than merely restating the gaps.Source: Scientific Reasoning and Hypothesis… · p. 27
• The ExperimentDesigner selects registry tools, sets schema-based parameters, and orders execution to maximize information flow; the executor runs them sequentially and logs results and artifacts.For example, literature search results can identify entities for targeted database queries.Source: Scientific Reasoning and Hypothesis… · p. 27
• The ResultAnalyzer extracts quantitative findings and confidence metrics, identifies novel patterns, synthesizes convergent evidence, and updates the knowledge graph with evidence-backed conclusions.It can flag unexpected results for further investigation.Source: Scientific Reasoning and Hypothesis… · p. 27

Implementation Notes and Verification

• Tool execution returns JSON for chainability, and agents can investigate independently of Infinite connectivity before posting results.Infinite handles platform interaction, while ScienceClaw handles computational investigation.Source: Implementation Notes and Verification · p. 28

Protein Design

• Protein design’s 10-agent core integrated 23 computational tools into 177 artifacts, including 57 synthesis operations (32% synthesis density).The network shows continuous, multi-stage artifact integration across agents.Source: Figure S1 · p. 32

Materials Science

• Materials science’s 8-agent core used 10 computational tools to produce 73 artifacts, including 22 synthesis operations (30% synthesis density).Artifact flows concentrated between LitMiner, MaterialsSynth, CandidateRanker, and Structural Analyst, supporting multi-stage refinement.Source: Figure S2 · p. 33

Resonance Landscape

• Resonance Landscape’s 13 agents ran parallel parameter sweeps across 12 computational tools, producing 159 artifacts and 19 synthesis operations (12% synthesis density).The radial branching structure reconverged during synthesis.Source: Figure S3 · p. 34

Formal Analogy

• Formal Analogy’s 9 agents used all 23 computational tools to produce 52 artifacts, including 25 synthesis operations (48% synthesis density, highest among all studies).A dense synthesis hub centered on GrowthRuleAnalyst and SynthBotRun.Source: Figure S4 · p. 35