MCP
MCP exposes tools and context to agents, making capabilities portable across clients and runtimes.
Concept
Agent orchestration and Human-Agent Collaboration solve different layers.
Orchestration answers how work should be decomposed, routed, sequenced, retried, or parallelized. Human-Agent Collaboration answers who is participating, what context is visible, what the user approved, what remains local, and how the durable conversation stays coherent across agents.
| Dimension | Agent orchestration | Human-Agent Collaboration Client |
|---|---|---|
| Primary question | How should work execute? | How should people and visible agents collaborate under explicit authority boundaries? |
| Core unit | Task, graph node, tool call, worker, run | Conversation, participant, membership, disclosure, approval, local effect |
| Human role | Initiator, supervisor, collaborator, or approver of a workflow | Persistent participant and authority for protected disclosure and local effects |
| State | Run state, checkpoints, retries, scheduling state | Durable conversation history, membership, context visibility, trust and approval state |
| Agent diversity | Can abstract workers behind execution contracts | Preserves useful native semantics across heterogeneous agent interfaces |
| Protocols | Can use MCP, RPC, queues, process APIs, or internal graphs | Composes ACP, MCP, RPC, HTTP/SSE, App Server, CLI, and federation protocols into the human-facing experience |
| Failure question | Should the task retry, branch, cancel, or fall back? | What evidence should the user see, who retains authority, and what context or effect remains valid? |
They compose
A collaboration client can contain orchestration as one of its capabilities.
LicoUp’s Adaptive Flywheel is an orchestration-oriented capability inside the broader conversation product. It can model immutable graph revisions, route selection, bindings, authorization, and run reduction.
The Canonical Conversation provides the durable human-facing center: memberships, structured events, conversation history, and private runtime bindings. Workflows can run inside that conversation while remaining independently inspectable.
Why the distinction matters
Clear layers make complex agent systems easier to reason about.
Models, agent runtimes, orchestration engines, relays, and protocols each solve different parts of the system. A Human-Agent Collaboration Client gives the user one place where participation, conversation state, approvals, context, trust, and effects come together.
That separation lets the underlying infrastructure evolve without constantly redefining the human-facing collaboration model.
Protocols and execution
Protocols connect the collaboration layer.
ACP
ACP provides a structured agent-client process boundary that a collaboration client can integrate into the wider conversation experience.
Lico Arc
Lico Arc is LicoLand’s protocol layer for endpoint communication and federation, giving LicoUp a protocol foundation for cross-endpoint collaboration.
Orchestration graphs
Graphs encode execution strategy while the collaboration client keeps the surrounding participants, context, approvals, and visible lifecycle coherent.
LicoUp today and next
The product already separates conversation and workflow authority in source.
The Canonical Conversation backend and Assistant workflow/Subagent MCP capabilities are implemented in source, giving LicoUp separate foundations for durable collaboration state and temporary workflow execution.
The roadmap extends the same separation into federation, human messaging, recovery, notary, and multi-device history, with Lico Arc providing the evolving endpoint protocol layer.
Continue