Sovereign agent nodes first. Framework details second.

---

Plaited is a framework for sovereign agent nodes.

Each node is owned by one user or organization, keeps its modules internal, records its own memory and provenance, and collaborates with other nodes through A2A instead of a central platform. The long-term goal is a modnet: a network of sovereign nodes exchanging tasks, services, artifacts, and controlled data without giving up local authority.

Under that model, Plaited provides the technical substrate:

• a BP-first runtime
• an agent loop with governance and evaluation seams
• generative UI primitives
• A2A and modnet building blocks
• git + .memory/-based provenance
• dev autoresearch infrastructure for improving the framework itself

Layer 1: What Plaited Is Building

Sovereign Nodes

A Plaited node is:

• one agent serving one user or organization
• one local workspace with its own constitution, memory, and modules
• one treaty boundary where all external coordination goes through A2A and PM mediation

Modules stay internal to the node. Other nodes do not directly access them. They request services, receive artifacts, or negotiate bounded work through A2A.

Modnets

A modnet is a network of sovereign nodes:

• nodes communicate peer-to-peer
• A2A is the external exchange boundary
• git bundles and other artifacts can later be exchanged through governed A2A flows
• enterprise and shared-service nodes are first-class future targets, not special cases

Native Model Direction

The framework and harness are being built first.

After that, a separate native-model lane focuses on making Falcon or another Plaited-native model strong at:

• generating modules end-to-end
• generating UI through Plaited's controller/generative UI model
• operating inside Plaited's BP/PM/MSS ontology

Layer 2: How Plaited Works

Behavioral Programming First

Plaited's core runtime is behavioral-program shaped:

• symbolic constraints and coordination are expressed as behavioral programs
• the neural model proposes work
• the symbolic layer governs, blocks, coordinates, and records it

This keeps:

• safety logic stable across model changes
• orchestration observable
• provenance and distillation tractable

Runtime Taxonomy

Current runtime work is converging on:

• MSS object — structural description only
• artifact — concrete implementation asset
• behavioral actor — BP runtime edge around an artifact
• sub-agent — isolated behavioral actor with fresh context/runtime
• team — coordinated set of sub-agents
• PM — sovereign coordinator and treaty authority

Agent Loop

The agent loop is a framework primitive, not a black box product feature. It provides:

• gating
• simulation/evaluation seams
• governance and constitution enforcement
• memory and snapshot handling
• local and A2A execution surfaces

Exports today include:

• agent.ts
• behavioral.ts
• modnet.ts
• runtime.ts
• ui.ts

Generative UI

Generative UI remains an important capability, but it now sits inside the broader node architecture.

The intent is not “AI-generated screens” in isolation. It is:

• UI generated by sovereign nodes
• shaped by local modules and memory
• constrained by behavioral programs and constitutions
• delivered through Plaited's controller/generative UI runtime

Provenance and Memory

Plaited separates two uses of memory:

• this framework repo's top-level .memory/
  • local dev/eval/autoresearch traces
  • ignored from git
• future node/module .memory/
  • part of the node's operational record
  • expected to be committed as provenance

Git is not just deployment plumbing here. It is part of the accountability model.

Repository Map

• src/ — framework code that ships with the node
• scripts/ — eval, adapters, graders, dev autoresearch infrastructure
• skills/ — implementation patterns, references, operational tooling
• docs/ — design rationale and active architecture docs
• dev-research/ — bounded research programs for improving Plaited itself

Current research lanes:

• program.md — framework/runtime/autoresearch lane
• program.md — native-model/Falcon lane
• program.md — inter-node/modnet lane

Development

Requirements

• Bun >= v1.2.9
• git
• for judged autoresearch:
  • Anthropic access
  • Varlock + 1Password setup via .env.schema

Useful Commands

# Typecheck
bun --bun tsc --noEmit

# Full test suite
bun test src/ skills/ scripts/

# One bounded research run
bun run research -- ./dev-research/runtime-taxonomy/slice-1.md --max-attempts 1

# Longer unattended batch
bun run research:overnight -- ./dev-research/runtime-taxonomy/slice-2.md

Agent-Facing CLI

The repo ships an agent-facing CLI toolbox:

• plaited read-file
• plaited write-file
• plaited edit-file
• plaited list-files
• plaited bash
• plaited validate-skill
• plaited validate-thread
• plaited ingest-goal
• plaited ingest-skill
• plaited ingest-rules
• plaited discover-skills
• plaited typescript-lsp

See:

• cli.ts

Read Next

• ARCHITECTURE.md
• MODNET-IMPLEMENTATION.md
• HYPERGRAPH-MEMORY.md
• CONSTITUTION.md

License

ISC