Give the assistant a role.
Not a blank check.
Build an agent with your models, skills, and tools, with MCIR memory already part of theo. An E.V.I. (Embedded Virtual Intelligence) brings that combination into your product, so relevant context can continue beyond one conversation.
Release desk
- Role
- Prepare a release summary from approved notes.
- Boundary
- Never mark a release approved or deploy it.
- Return
- A draft, its source references, and unresolved questions.
01four different responsibilities
Persona, skill, tool, memory. Keep the distinction.
- Persona behavior
- Sets the assistant’s role, tone, and instructions. “Be concise and identify missing evidence” guides behavior. It does not create a permission boundary or guarantee that an answer follows the instruction.
- Skill domain package
- Packages prompt extensions, tool definitions, permissions, model preferences, triggers, and reference knowledge. Installed skills and per-request skill slugs are separate ways to make that context available. explore the skill development loop
- Tool callable action
- Describes an operation with a name, purpose, and input schema. The model can decide to request it. Your integration must provide appropriate execution, authorization, validation, and failure handling.
- MCIR memory built-in continuity
- Brings relevant history, decisions, and preferences to the task within the authenticated owner’s scope. An agent’s persona or a shared skill does not grant access to another user’s memory.
02E.V.I. Canvas
Compose the skill before you publish it.
E.V.I. Canvas is the visual skill builder. A connected, directed acyclic graph is compiled into a skill manifest and workflow steps. This is a different job from choosing the request path in an orchestrator graph.
The documented Canvas also supports Vision and Condition nodes. A canvas requires one Input and at least one reachable Output. Compilation checks structure; it does not prove the assistant’s answers or tool permissions are correct.
03from a brief to a reviewed skill
A short development loop, with a real stop before action.
- 01
Design the boundary
Write what the assistant should do, what context it needs, and what it must leave to a person. Prefer a small read-only starting scope. - 02
Compile the structure
Connect the canvas, save it, and compile. Resolve cycles, unreachable nodes, and missing prompt information before testing behavior. - 03
Test with synthetic data
The documented Canvas sandbox uses mock data and no real tool calls. Try missing sources, ambiguous requests, and tool failures, not only the happy path. - 04
Choose visibility deliberately
Private, organization, and public marketplace visibility have different audiences. Publishing enters the skill review process; it is not an automatic deployment approval.
two canvases, different purposes
A skill graph is not the orchestrator.
E.V.I. Canvas builds a skill.
Its Input, Prompt, Tool, Knowledge, Vision, Condition, Model, and Output nodes describe a package of behavior and capabilities. The output of compilation is the skill structure.
Orchestrator graphs coordinate requests.
Their entry, classifier, router, model, guardrail, and description nodes describe the route through a configured request. Skills attach to model nodes; a tool or knowledge source is not an interchangeable orchestrator node. The marketing demo adds an Output endpoint for its application handoff, separate from the current product editor.
see the platform handoffsDoes a persona enforce a policy?
No. Persona instructions influence behavior. Gateway Guardrails require an explicit policy binding, and application authorization must still protect data and actions.
Does every declared tool run?
No. Tool definitions make operations available; descriptions help the model decide whether to request them. Availability, permissions, and the execution path determine what can actually happen.
Should every skill be installed?
No. Start with the context and actions needed for a specific task. Installed and per-request skills can stack, so review the combined behavior and access rather than considering each in isolation.
Compose the agent. Keep the context.
Shape a skill in E.V.I. Canvas, choose the models and tools, then deploy into your application with owner-scoped MCIR memory behind the work.