Labtune: Treating Lab Organization and Information Sharing as Experimental Variables in LLM-Agent Discovery
Abstract
De-siloing can create groupthink; silos can prevent knowledge sharing and reduce research efficiency. The same tension appears in research systems built with large language model (LLM) agents, yet their organization is usually fixed by design. I propose treating laboratory structure and information sharing as experimental variables. I introduce Labtune, an open framework for bounded research workflows with user-selected laboratory organization, communication policy, model configuration and budget. Users can choose among six editable designs, from multidisciplinary silos and transdisciplinary teams to co-leaders, hub-and-spoke structures, flat co-labs and a single scientist. Information sharing can independently vary from none to high detail, with an optional channel for conditional abstractions called wisdom. Sharing is enforced by the controller at phase barriers rather than through unrestricted agent chat, while delegation does not create additional budget or permissions. Development evidence is separated from protected confirmation, and an optional multi-lab layer allows laboratories with different structures to exchange filtered information. Labtune is an idea and system-description paper, not a benchmark result. Its contribution is to make laboratory organization and information sharing configurable and measurable under matched budgets. I do not claim that one design or level of sharing is best; the aim is to create a concrete experimental space in which those differences can be tested.
Figures & tables
| Design | Organization and personas | Phases | Suggested |
|---|---|---|---|
| A Multidisciplinary silos | Formal, empirical and implementation specialists pursue separate tracks | Independent framing, parallel episodes, independent verification, optional handoff | low + wisdom |
| B Transdisciplinary integration | Structural modeler, experimentalist, systems designer, plus a method integrator | Independent framing, translation workshop, parallel tests, framework revision | high + wisdom |
| C Complementary co-leaders | A mechanism lead and a measurement lead, each with a team | Independent positions, parallel episodes, paired challenge, discriminating test | intermediate + wisdom |
| D Hub-and-spoke matrix | A research lead coordinating creative, technical and measurement tracks | Track charters, parallel episodes, hub review, scoped handoffs | intermediate + wisdom |
| E Flat / holacratic co-lab | Exploratory, builder and critical peers; rotating facilitator | Independent proposals, roundtable, task adoption, evidence review | intermediate + wisdom |
| F Single scientist | One scientist with low, medium or persistent effort | Investigate, then review and refine | none |
| Mode | Admitted content | Intended trade-off |
|---|---|---|
| none | No peer research content; operational scheduling only | Baseline that preserves independent search |
| low | Notable dead ends only, each backed by a valid development receipt; no positive results, rankings or recipes | Avoid repeating costly failures without steering toward a solution |
| intermediate | Compact findings, broad method descriptions, constraints and trade-offs; no exact code or tuning recipes | Share useful direction while limiting copying |
| high | Experiment summaries, measurements, methods, parameters, permitted artifact references | Deliberate reuse and recombination |
| wisdom_only | Conditional abstractions only (lesson, conditions, limits) | General lessons with reduced solution steering |
| Layer | Implemented | Not implemented / limits |
|---|---|---|
| Structural core | Six designs, schemas, strict YAML, phase DAGs, configuration identities | — |
| Study runner | Explicit provider calls, bounded concurrency and delegation, JSON correction, filtered Markdown sharing, receipts | Live mid-run reconfiguration; durable scheduler recovery |
| Multi-lab | 2–32 labs, frozen exchanges, conferences, symposia, sequences, group budgets | No pooled quality score or winner |
| Statistics | Clopper–Pearson intervals; Beta–binomial posteriors; reviewed priors | Other likelihoods are adapter-specific |
| Domain tools | Registered adapters; sandboxed synthetic memory-policy task | Candidate execution only on macOS sandbox |
| Visualization | Local read-only setup visualizer and runtime observer; offline HTML snapshots | Shows metadata, not prompts or results |
Multi-Agent LLM Systems
Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.
See all 1,100 papers