01
Intent-First
Coordination starts from a goal, not from an endpoint or schema. If it moves things toward the objective, it's valid.
Say what, not how.An open protocol for multi-agent coordination. No pre-wired APIs, no shared schemas, no glue code. Agents state intents in natural language — discovery, negotiation, execution, and cleanup happen at runtime. Ships with a full admin dashboard, REST API, and CLI.
An AI agent can reason, adapt, and pursue complex goals. But the moment it needs to talk to another system, it's stuck calling rigid endpoints with exact JSON schemas. The mismatch is the bottleneck.
Every connector you write, every schema you sync, every breaking change you manage — that's hidden technical debt that compounds with each new agent or system.
"Deliver this container within 200 km of the closed port, optimizing for cost and time."
That's it. The requesting agent states its goal. The runtime discovers who can help, negotiates terms, composes an execution plan, runs it, and dissolves everything after. No pre-wired endpoints. No shared schemas. No leftover coupling.
Agent states what it needs in natural language
Runtime finds agents whose capabilities match semantically
Agents exchange offers, agree on terms and constraints
Coordinated work runs under a dynamically built graph
Everything is cleaned up. Zero residual coupling.
Coordination starts from a goal, not from an endpoint or schema. If it moves things toward the objective, it's valid.
Say what, not how.Every interface lives only for the duration of the task. When it's done, schemas, auth scopes, and bindings are all invalidated.
Zero technical debt, by construction.Agents that have never met before can collaborate. Terms are agreed at runtime through semantic negotiation.
No prior integration required.A single intent can be fulfilled by combining capabilities from multiple agents. The runtime composes them into a single execution graph.
Multi-agent coordination, not orchestration.Authorization evaluates what is being attempted and why — not just who is asking. Governance acts on purpose, not identity.
IBAC over RBAC.When something fails, agents don't crash — they renegotiate. Constraints are revised, alternatives are proposed, context is preserved.
Errors invite clarification, not collapse.The protocol is real. This is the running implementation — with a full admin dashboard, REST API, and CLI.
# Install
git clone https://github.com/draiven-io/agentic-bus.git
cd agentic-bus
pip install -e ".[dev]"
agbus install # interactive setup wizard → creates .env
# Start the coordinator (WebSocket :8765 + Admin API :8766)
agbus serve
# Start the admin dashboard (http://localhost:3000)
cd ui && npm install && npm run dev
# Run a sample agent
python -m app.agents.examples.logistics_agent.agentSemantic matching finds the right agents for any intent — no catalog lookups or hardcoded routing.
Execution plans are synthesized as LangGraph DAGs at runtime, not defined ahead of time.
Intent-Based Access Control with a full rule editor. Policies evaluate purpose and context, not just identity.
Full-featured Next.js 16 UI — manage agents, inspect intents in real-time, configure IBAC rules, tenants, and users.
FastAPI admin API with 40+ endpoints, Swagger docs, and full CRUD for agents, tenants, users, IBAC rules, and LLM configs.
First-class tenant isolation. Create tenants, assign agents and users, scope access per organization.
Create and manage agents directly from the coordinator with CrewAI tool integration. No external deployment needed.
Real-time bidirectional communication. Agents connect, authenticate (OIDC), and coordinate over persistent channels.
OpenTelemetry tracing, audit logging for every admin action, and a searchable audit trail in the dashboard.
OpenAI, Azure, Anthropic, Google Gemini, Ollama — manage multiple configurations and switch the active provider from the UI.
18 test files covering every subsystem — protocol, negotiation, graph, IBAC, persistence, admin API, tenants, and more.
Full administration from the terminal — agent lifecycle, database management, interactive setup wizard, and configuration.
agbus install # Interactive setup wizard
agbus serve # Start coordinator (WS + REST API)
agbus db init # Create / migrate database
agbus config show # Display current config
agbus agent list # List registered agents
agbus agent approve <id> # Approve a pending enrollment
agbus agent reject <id> # Reject an enrollment
agbus agent revoke <id> # Revoke an approved agent
agbus agent create # Create a managed agent (interactive)
agbus agent activate <id> # Activate a managed agent
agbus agent add-capability <id> # Add capability to an agent
agbus agent tools # List available CrewAI tools"What tools are available to me?"
"What should happen, and who can help?"
MCP handles capability exposure — what an agent can do. LIP handles coordination — how agents negotiate to fulfill complex intents across boundaries. An MCP-enabled agent can participate in LIP workflows out of the box.
Reference implementation. Coordination runtime with LangGraph, IBAC, admin dashboard (Next.js), REST API (FastAPI), multi-tenancy, and full CLI.
View on GitHub →The core protocol for intent expression, semantic negotiation, and ephemeral contract formation.
View specification →Intent-Based Access Control. Governance that evaluates purpose and context, not just identity.
View specification →"Liquid Interfaces: A Dynamic Ontology for the Interoperability of Autonomous Systems" — arXiv 2601.21993
Read on arXiv →Agentic Bus is MIT licensed. The protocol spec is Apache 2.0. Everything is open.