Skip to content
Contact
☰
Color scheme: Dark.
Back to research
RESEARCH SPACE E

Distributed, Agentic & Collective Intelligence

How does coordinated intelligence arise when perception, knowledge, decision and action are distributed across several specialized artificial units?

What this space studies

How several specialized agents, models and technical systems work together. EIL uses running agent systems as a research environment, because delegation, authority and failure can be observed in the logs. Concretely: specialization, handoffs, repair and hidden centralization. It also asks when observed coordination follows from explicit orchestration and when it might arise only from the interaction of several components. The space is not a catalog of agent products.

Central questions

  • How do agents divide tasks and responsibility?
  • When does hidden centralization arise?
  • How do distributed systems handle failure?

Research object

Distributed, agentic and collective intelligence in running multi-agent systems: specialization, coordination, state and failure.

Designed versus emergent coordination

The central question is when observed behavior follows from explicit orchestration and when it may arise only from the interaction of several components. If the behavior can be fully derived from orchestration rules and handoff protocols, it counts as designed coordination. If it cannot be derived, emergent coordination remains a hypothesis until protocols across several runs and deliberate variation of single components support it.

Systems as study objects

The named systems are research setups, not evidence. JARVIS serves as a study object for orchestration and coordination, without evidence that its orchestration is effective. MIRROR studies the silent loss of a previously correct state, JANUS authority and provenance at the handoff points, and DAEDALUS distributed reasoning with a checkpoint. DELTA and KOPERNIKUS serve to study specialization, role formation, delegation and handoffs, LYNX specialization and failure behavior, ROBERT delegation in a narrowly bounded task field, ATLAS the merging of agent reports. NYX studies failure and recovery under attack in its own test environments, ARGUS state and continuity over time.

Research programs

1 program of its own, 1 more touch this space.

Further programs that touch this space (1)

Core phenomena

  • Agentic systems
  • Specialization
  • Coordination
  • Collective intelligence
  • Failure
  • Externalization of human functions (central)

Core phenomena

Agentic systems

  • agents
  • sub-agents
  • lead agents
  • degrees of autonomy
  • roles
  • tools
  • memory
  • state

Specialization

  • functional specialization
  • emergent specialization
  • designed roles
  • dynamic roles

Coordination

  • delegation
  • orchestration
  • handoffs
  • authority
  • scheduling
  • shared state
  • designed coordination
  • emergent coordination

Collective intelligence

  • collective reasoning
  • multi-perspective analysis
  • division of labour
  • aggregation
  • conflict resolution
  • distributed cognition
  • self-organization
  • feedback between agent and environment
  • unintended collective behavior

Failure

  • distributed failure
  • hidden centralization
  • state divergence
  • handoff failure
  • role drift
  • memory inconsistency
  • recovery
  • shared and conflicting states
  • coordination failure
  • collective recovery

Externalization of human functionscentral

  • agent identity and continuity
  • artificial cognitive organization
  • human cognitive functions as specialized artificial functions

Keywords (6)

  • agent intelligence
  • retry
  • long-running intelligent systems
  • cognitive function externalization into artificial systems
  • role drift and state divergence
  • coordination failure and collective recovery

Further research questions of this space (8)

  • What cognitive functions externalize?
  • Under which conditions do they externalize?
  • How stable are they?
  • How are they represented: as method, model, memory structure, agent or computational system?
  • How do they change after externalization?
  • What happens when they are reintegrated into future interaction?
  • Which cognitive operations become externalized?
  • How do externalized structures change subsequent human behavior?

3 questions of this space's programs are with the program details →

How this space is investigated

Observation and relation

Distributed intelligence is studied on running agent systems.

Observation and relation

What is specifically observed

Handoffs, failures, retries and repair are observed from logs. Specialization or hidden centralization are inferred or hypothesized until the logs support them across several runs. Statements about what an agent does internally count as unknown as long as only its behavior is available. The central distinction is between designed coordination, which follows from explicit orchestration, and emergent coordination, which may arise only from the interaction of several components. Which of the two applies is a hypothesis until logs and variation of components tell them apart. Collective intelligence and autonomy are research questions here, not established properties.

Relation to the other spaces

This space is the point in the research loop where an externalized thinking operation continues as a role, method or agent inside an artificial system. The space studies what happens to it once it gains memory, reasoning and orchestration.

Intelligent & Computational Research Systems

Designing the computational and intelligence infrastructure needed for stateful, multi-step and multi-agent research.

Central question

What architecture keeps relevant state, evidence and context accessible across long, multi-step, multi-agent research processes instead of letting it get lost in a single session's working context?

Intelligent & Computational Research Systems

Scope
  • Intelligence Design
  • Agent Intelligence
  • Multi-Agent Systems
  • EIL Kernel
  • Research runtimes
  • Data structures
  • Scientific visualization
  • Experimental infrastructure
  • Agent registration & role handoffs
  • Emergent collective phenomena
Approach

Identify recurring analytical operations and translate them into specialized agent roles with clear inputs, outputs and handoffs, hold state and evidence centrally rather than per agent, and build infrastructure only after genuine, repeated research demand.

Every observed recurrence is first tracked with its own identifier, evidence base and counter-evidence in a central research map before it is assigned to a publication.

No figure may imply stronger causal or substantive knowledge than the evidence supports, and inferred elements are visually distinguished from observed ones.

The mission lifecycle stays observable, from classification through plan generation, role handoffs and interim review to acceptance.

Unintended collective phenomena such as spontaneous specialization, hidden centralization or recurring coordination shortcuts are an observation object in their own right.

Relationship

The infrastructure layer behind the EIL Kernel, the Multi-Agent Research Environment and the Research Knowledge Graph: what makes the other seven spaces operationally executable.

Agents as externalized functions

Planned The investigation is functional. It asks which recurring human operations were formalized as artificial roles, how these roles were recombined, and what changed measurably afterwards. The list of required evidence is with the methods. This must not be reduced to agents having simply been programmed or to agents being only personas.

Required evidence in the methods →

Methods, systems and axes

Methods (4)

All methods of this space

Methods and analysis procedures (4)

See all methods →

Systems (22)

All systems of this space

Research systems (2)

Multi-agent systems (11)

Shared technical foundation (2)

System state, memory, security & integrity (7)

See all systems →

Cross-cutting research axes (5)

How the axes act in this space

Cross-cutting condition Ethics / Human Agency / Human Rights →

Evidence & limits

Collective intelligence is not demonstrated

Collective intelligence and autonomy are research questions in this space. The systems run as research setups. For JARVIS the effectiveness of orchestration is not evidenced, for MIRROR no detection rates are published, for ROBERT no effect on response rates is evidenced. A coordinated overall performance counts as collective only once it can be told apart whether it stems from individual agents, central control or distributed interaction.

Boundary of this space

Coordinated performance of a whole system is not automatically collective intelligence. It has to remain distinguishable what arises from individual agents, central control or truly distributed interaction.

How to read status and origin

Every statement carries its status (observed, inferred, hypothesis, planned or unknown), no program is externally validated, and origin statements about systems are candidates.

Evidence register → Origin of the systems →

Topics and publications

working formulation Formulated from the research program, not yet a fixed wording. Exception: 01 Emergent Intelligence.

Topics with this space as main space (2)

Further topics that touch this space (1)

Guiding questions of the topics

  • 05 Multi-Agent and Collective Intelligence

    When does orchestration produce collective capabilities?

  • 06 Memory, State and Identity

    What should a system keep over time, and when does memory produce false continuity?

  • 01 Emergent Intelligence

    When does interaction stop being merely interaction and begin producing persistent intelligent structure?

Publications (4)

Publications (4)