Skip to content

Research

The Research section documents the active agentic-systems research lineage that Emergence and Fabric draw on, organized so that every developer-facing pattern in Emergence and Fabric can be traced back to the published research that motivates it. Each subpage collects the canonical systems for one platform layer, summarizes their mechanics in two-to-four sentences, and provides explicit selection criteria for picking one over another in a given Unitt fabric workload.

This section is not a literature review for its own sake. It is the platform's working bibliography, used by developers when choosing a runtime pattern, by operators when validating an architectural decision, and by the WorldSim evolutionary loop when proposing mutations. Each subpage closes with a selection-criteria table that the platform can read directly and use as the prior distribution over runtime configurations.

Emergence Sections

  • Memory Systems; MemGPT / Letta, A-MEM, Mem0, HippoRAG, Generative Agents, CoALA cognitive architecture, hybrid stores.
  • Runtime Systems; ReAct, Reflexion, ReWOO, Plan-and-Execute, Tree-of-Thoughts, Graph-of-Thoughts, LATS, Self-Refine, CodeAct, Voyager, reasoning-native loops, agentic-OS runtimes.
  • Context & State; prompt caching, layered window, tiered memory, sliding window with attention sinks, compaction, hybrid retrieval, contextual retrieval, long-context, isolation, checkpointing, context engineering.
  • Subagents; Claude Code subagents, Anthropic multi-agent research system, OpenAI Agents SDK, LangGraph supervisor / hierarchical teams, CrewAI, AutoGen / AG2, MetaGPT SOP, AgentVerse / ChatDev, debate and consensus.
  • World Simulation; Generative Agents, Voyager, AgentBench, AgentBoard, WebArena, OSWorld, ALFWorld, SWE-bench Verified, τ-bench, GAIA, BrowseComp, MLE-bench, MiroFish / OASIS social-sim.

Fabric Sections

  • Fabric Setup; multi-agent governance, archetypes, role design, model tier mix, authorization, identity, budgets, observability.
  • Fabric Data; MCP, enterprise bindings, RAG, streaming, schema enforcement, egress guardrails, lineage, multi-tenant isolation.
  • Fabric Flow; staged flows, MAST taxonomy, loop / deadlock detection, circuit breakers, retry budgets, timeouts, escalation, sagas, validation gates.
  • Fabric Test; benchmarks, pass^k, cost-per-success, synthetic + real data, LLM-as-judge, replay regression, red-team, stress.
  • Fabric Publish; marketplaces, packaging standards, A2A, deployment targets, parameterization, versioning, licensing, telemetry, GitOps reuse.

Assembly Sections

  • Assembly Core; identity, rules, policies, governance, NIST AI RMF, OWASP Agentic Top 10, Cedar / OPA, PEP/PDP, audit, escalation, provenance.
  • Assembly Objectives; outcome-shaped objectives, Plan-and-Solve / Plan-and-Execute / HTN / MetaGPT decomposition, pass^k, constraints, validation.
  • Assembly Patterns; the six Anthropic canonical patterns, LangGraph / AG2 / CrewAI / LlamaIndex, branching, parallelization, refinement, HITL.
  • Assembly Connectors; MCP, OAuth 2.1 + PKCE + DPoP, Workload Identity Federation, vaults, validation states, MCP Gateway, tool poisoning defense.
  • Assembly Tools; description quality, strict-mode function calling, CodeAct, registries, tool poisoning, semver, default Unitt tool mapping.
  • Assembly Skills; Anthropic Agent Skills, progressive disclosure, CoALA procedural memory, Voyager skill library, skill-creator evals, skill safety, marketplaces.

How Selection Criteria Are Applied

Every Emergence layer exposes a selection variable that the platform can set per Unitt or per workflow stage. The Research subpages define the criteria the platform uses to pick a value for that variable. Where multiple criteria conflict, the platform applies a Pareto selection across outcome correctness, cost, safety, and latency before falling back to a domain-default. The selection criteria are also the substrate on which the WorldSim evolutionary loop mutates Unitt configurations across generations.

flowchart LR
    UC[Unitt Workload Profile] --> SEL[Selection Criteria]
    SEL --> M[Memory Choice]
    SEL --> S[System Choice]
    SEL --> ST[State Choice]
    SEL --> SU[Subunit Composition]
    SEL --> WS[WorldSim Environment]
    M & S & ST & SU & WS --> CFG[Resulting Configuration]
    CFG -. validates against .-> WSV[WorldSim Validation]
    WSV -. feedback .-> SEL

    classDef stage fill:#ffd541,stroke:#222021,color:#222021
    class UC,SEL,M,S,ST,SU,WS,CFG,WSV stage

Citation Style

Every Research subpage cites primary sources directly. Where a system has an arXiv paper, the paper is linked first. Where a system has an official engineering blog, repository, or documentation page, those are linked second. The platform treats published primary sources as load-bearing; if a referenced system updates or supersedes its design, the corresponding subpage is updated and the Emergence layer that depended on it is regenerated.