Teamknitt: Configuring, Observing and Intervening in LLM-Agent Teams through the Lens of Dynamical-Systems and Human-Team Theory
Abstract
Most frameworks for teams of large language model (LLM) agents fix the team's shape inside the software: who talks to whom, how much is passed along, and what a human can do once a run has started. Human-team research has long asked those questions: where team phenomena live (in members or in their interaction), how team-level properties arise, and what separates what a team does from how a team is. I propose using that theory as the organizing lens for configuring, observing and intervening in LLM-agent teams, without assuming its findings transfer. I introduce teamknitt, a framework and open reference implementation in which a user configures a team of 1-16 AI agents for task completion, solution generation, achieving a goal or a research mission; selects one of fourteen deterministic routing topologies or a custom directed graph; chooses a sharing profile that limits what may travel along permitted routes; and binds sources, representations, work dependencies, handoffs and an expertise directory. The recorded process can be observed through non-system views (member reports and behavior) and system views (distributed problem-solving, differentiated transactive memory, interactive coordination, team dynamics). Interventions are explicit and human-led: reviewed reconfiguration between runs, planned disturbances, and private feedback, pause, resume or stop applied at dependency-wave boundaries, with "queued", "applied", "assembled", "understood" and "improved" kept as separate claims. Emergent states such as trust and shared understanding are declared unmeasured. This is an idea and system-description paper with falsifiable hypotheses on topology, information flow and intervention timing, including whether the fractal scaling of communication output between members, and between members on different models, relates to task and team performance. No comparative result is reported.
Figures & tables
| Approach | Category | Main focus | Labtune / teamknitt emphasis |
|---|---|---|---|
| Tree / Graph of Thoughts [ 45 , 2 ] | Reasoning methods | Reasoning paths and dependencies between intermediate thoughts | Organization of agents and teams; reasoning methods can operate within individual members |
| Agent Laboratory [ 39 ] | AI research system | Specialized agents covering literature review, experimentation and reporting | Alternative lab designs, topologies and exchange rules as user configurations |
| AgentRxiv [ 38 ] | Research-sharing infrastructure | Shared reports so autonomous labs can build on each other’s findings | Configurable exchange within and across teams and labs: who shares, when, under which rules and word budgets |
| AI co-scientist [ 12 ] | AI research system | Hypothesis generation, review and evolution through specialists and tournaments | Exploration of different organizations and exchange styles rather than one fixed process |
| MacNet [ 36 ] | Multi-agent architecture | Graph-based collaboration; effects of topology and agent scaling | Topology combined with sharing profiles, bindings and user-led follow-up and intervention |
| GPTSwarm [ 47 ] | Agent graph optimization | Optimization of prompts and information-flow connections | Explicit, stable configurations for examining how organizational choices affect results |
| Preset | Exact routing |
|---|---|
| fully_connected | Every ordered pair of distinct members |
| hub_and_spoke | Member 1 exchanges two-way with every spoke; no spoke–spoke links |
| ring | Two-way neighbours with a closing loop |
| hierarchical | Two-way binary tree over member ids; no decision authority is granted |
| chain | Two-way adjacent ids, no closing loop |
| directed_chain | ; no implied replies |
| Profile | Admitted content | Intended trade-off |
|---|---|---|
| none | No peer content; independent contributions retained for user review | Baseline that preserves independent search; joint phases are operational only |
| low | Bounded member-reported blockers, labelled unverified | Avoid repeated dead ends without steering toward a solution |
| intermediate | Compact contribution summaries | Share direction while limiting copying |
| high | Detailed peer views and explicitly granted versioned notes | Deliberate reuse and recombination |
| wisdom_only | Conditional lessons with stated limits, not detailed peer output | General lessons with reduced solution steering |
| View | Inspects | Not measured or inferred |
|---|---|---|
| Non-system: member- and component-centred | ||
| Collaborative intelligence | Eight teammate-behaviour questions [ 44 ] , from an optional, separately run probe protocol | Capabilities of the configured team; an official benchmark score; psychological traits |
| Shared models & member reports | Explicit plans, agreement with a declared reference, report freshness, perceived alignment | Latent mental-model similarity or accuracy; understanding; correctness |
| System: interaction- and configuration-centred | ||
| Distributed problem-solving | Complementary inputs, representations, tools, expertise directory, transformations, dependencies, handoffs | Emergent cognitive capacity; goal achievement |
| Transactive memory | Differentiated (compilational) expertise assignments, holders and retrieval scaffolds | Credibility, specialization, retrieval success, system effectiveness |
| Timing / mechanism | Actual support | Boundary |
|---|---|---|
| Between runs: inputs and conditions | Reviewed changes to roles, task, context, criteria, sources, procedures, topology and sharing | Draft first; explicit apply with private backup under the run lock; no automatic staffing or model choice |
| Before a run: planned disturbance | actor_unavailable , representation_unavailable or link_dropout at a chosen wave for a chosen duration | A software condition for a future run, not weight training or a validated adaptive-training replication |
| During a run: private feedback | Bounded task context for chosen members’ future frames; newest text first within the representation word cap | No new tool, source or permission grants; not proof of delivery or understanding |
| During a run: pause / resume / stop | Acknowledged at the next dependency-wave boundary; current-wave work may finish | Pause keeps the lock, counters and wall deadline; resume resets nothing; stop is terminal and skips confirmation |
| After a run: outcome review | Private partial or complete deliverable, request receipts, separate outcome evidence | No automatic quality, psychological-state or intervention-effect score |
| Layer | Implemented | Not implemented / limits |
|---|---|---|
| Configuration | Four purposes; 1–16 members; 14 presets plus custom links; five sharing profiles; schema-3 bindings; drafts, backups, run lock | No automatic topology, staffing or model selection |
| Runtime | Dependency waves, scoped frames, immutable versioned notes, route-filtered views, barrier publication, handoffs, plan reports | Bounded text by default; live mid-run reconfiguration; durable crash recovery |
| Views | Non-system and system groups; live metadata monitor; private offline snapshots; optional probe attachment | Metadata only, never raw prompts, generated text or credentials |
| Observations | Local–global, perturbation, coupling and cross-level descriptions | Attractors, synchronization and fractals not estimated; no emergence or causal inference |
| Interventions | Private file queue; feedback, pause, resume, stop at wave boundaries; planned disturbances | Not available to advanced engine workflows implicitly; no retries or recovery |
| Statistics | Descriptive records; frequentist count and interval analysis where applicable | Bayesian analysis and prior elicitation removed; unsupported legacy choices fail before provider initialization |
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