Skip to content

Agent capabilities

Compare current Copilot and Designer MCP states, authority boundaries, and the evidence required before using each workflow.

Updated View as Markdown

Use this matrix before prompting a model or designing an agent workflow. The machine-readable source for the public release state is ai-capabilities.json.

Capability State What is allowed
Copilot Ask Controlled-release beta Read-only answers about the open chart
New conversation Released Start with fresh conversation context; chart is unchanged
Coding-agent docs and toolkit Released Help implement SeatLayer in a repository using canonical docs
Designer MCP Preview Work through OAuth and one authorized chart scope
Copilot Edit Off Do not expose or claim chart modifications
Multistep chart creation Off Do not claim prompt-to-published-chart creation
Photo/reference creation Off Do not infer authoritative inventory from an image
AI-led 3D improvement Off Do not claim automatic height or 3D improvement

Authority does not move into the model

Concern Authority
Authentication, role, workspace, and chart scope SeatLayer API
User confirmation and publish permission SeatLayer API and human reviewer
Tool availability and spend limits SeatLayer product configuration
Chart writes and canonical revision SeatLayer chart service
Idempotency, retry, and conflict handling SeatLayer execution records
Validation and rendered evidence SeatLayer validators and renderers
Usage, billing, and terminal state Canonical SeatLayer records
Narrative explanation and proposed plan Copilot or connected model

A model cannot authorize itself, invent a user confirmation, mark an action complete, or publish by returning a persuasive message.

Release proof for write capabilities

Before Edit or Create can move out of off, the workflow needs:

  1. exactly-once or safely idempotent execution across retries and resumed runs;
  2. revision conflict detection and a full-scope rollback plan;
  3. renewed authorization and kill-switch checks when a run resumes;
  4. deterministic completion validation against the canonical chart;
  5. visible entitlement, token/model, rendering, and workspace spend controls;
  6. privacy rules for prompts, references, traces, evidence, and retention; and
  7. repeated real-user validation across representative chart types.

AI-led 3D also needs actual WebGL before/after evidence from the same chart revision on desktop and mobile, plus a complete 2D fallback.

Continue with Copilot overview, connect Codex, or Designer MCP.

Navigation

Type to search…

↑↓ navigate↵ selectEsc close