Converge.zone

Converge is a correctness-first, context-driven multi-agent runtime built in Rust.

Agents collaborate through shared context, not by calling each other. The engine runs agents repeatedly until a fixed point is reached — convergence is explicit and observable.

Design Principles

1. Agents suggest, engines decide. ProposedFact is not Fact. Every proposal carries confidence and provenance. The promotion gate validates before promoting.
2. Context is the API. Agents communicate through shared context, never by calling each other.
3. Append-only truth. Facts are never mutated; corrections are new facts.
4. Convergence is observable. CriterionEvaluator checks success conditions. CriterionResult::Blocked triggers honest stopping with HumanInterventionRequired.
5. Truths are typed. TruthDefinition declares jobs, policies, and invariants with success criteria, constraints, and pack participation — not string descriptions.
6. Consequential transitions are policy-gated. Real business flow actions project into a neutral FlowGateAuthorizer contract, with Cedar as the default implementation.
7. Safety by construction. unsafe_code = "forbid" everywhere. Type-state enforcement on proposals → facts.

Quick Start

git clone https://github.com/Reflective-Lab/converge.git
cd converge

just build-quick    # build (fast iteration)
just test           # run tests
just lint           # format + clippy

Examples

just example hello-convergence   # engine loop, agents, facts
just example custom-agent        # implement the Suggestor trait
just example meeting-scheduler   # domain pack with constraints
just example custom-provider     # implement an LLM provider

See examples/README.md for the full list.

Governed Flow Gates

Converge now has a default gate path for consequential flow transitions:

• converge-core defines the neutral FlowGateAuthorizer contract
• converge-policy implements that contract with Cedar
• converge-domain applies it to real pack transitions like invoice issuance, period close, and contract execution

The same shape is exercised in the governed examples:

• expense-approval — finance commit escalation and approval
• vendor-selection — procurement commit escalation and approval

Architecture

crates/
├── pack/          # Canonical pack authoring contract
├── provider-api/  # Canonical provider capability contract
├── model/         # Curated semantic model surface
├── kernel/        # Curated in-process embedding API
├── protocol/      # Generated wire contract (converge.v1)
├── client/        # Canonical remote Rust SDK
├── core/          # Convergence engine (implementation)
├── traits/        # Deprecated compatibility facade
├── provider/      # LLM backends (Anthropic, OpenAI, Gemini, Ollama, ...)
├── domain/        # Domain packs (scheduling, routing, drafting, governed flows)
├── experience/    # Event-sourced audit store
├── knowledge/     # Vector knowledge base
├── mcp/           # Model Context Protocol (client + server)
├── optimization/  # Constraint solving (OR-Tools)
├── storage/       # Object storage abstraction
├── policy/        # Cedar policy engine and default flow gate authorizer
├── auth/          # Authentication, authorization, and cryptography
├── consensus/     # Raft consensus adapter for distributed Converge
├── ledger/        # Append-only context ledger
├── nats/          # NATS messaging adapter
├── observability/ # Audit, telemetry, and metrics
├── llm/           # Local LLM inference (Burn)
├── analytics/     # ML/analytics agents
├── tool/          # Development toolchain (Gherkin, JTBD)
├── runtime/       # HTTP/gRPC execution service
├── remote/        # Compatibility CLI on top of client + protocol
└── application/   # Reference distribution
examples/
├── hello-convergence/   # Minimal convergence loop
├── custom-agent/        # Implement the Suggestor trait
├── meeting-scheduler/   # Domain pack with constraints
├── custom-provider/     # LLM provider adapter
├── vendor-selection/    # Multi-criteria vendor selection with Cedar commit gating
├── expense-approval/    # Governed expense approval via default flow gate contract
├── loan-application/    # Loan application processing
├── local-inference/     # Local inference on Apple Silicon
└── gemma-inference/     # Gemma GGUF inference via llama.cpp

Canonical Public Crates

Six crates form the supported external API (see ADR-001):

Configuration

CONVERGE_LLM_BACKEND=ndarray
CONVERGE_LLM_MODEL=llama3
CONVERGE_STORAGE_BACKEND=lancedb
CONVERGE_STORAGE_PATH=./data
RUST_LOG=info

Documentation

• kb/Architecture/System Overview.md — system design and dependency graph
• kb/Architecture/API Surfaces.md — canonical public contracts
• DEVELOPMENT.md — setup, build, git workflow (worktrees, jj)
• CONTRIBUTING.md — contribution guidelines
• SECURITY.md — vulnerability reporting and security baseline
• crates/provider/.env.example — provider API key and endpoint catalog
• examples/ — runnable examples

License

MIT — see LICENSE.

Copyright 2024-2026 Reflective Labs

Kenneth Pernyer — kenneth@reflective.se

Project: github.com/Reflective-Lab/converge