Analysis
Figma added its agent to Dev Mode on Nov. 19, giving developers another way to inspect designs, compare versions, request code snippets and obtain asset download links. For design leaders, the useful question is not whether an agent can discuss a screen. It is which actions the person asking the question can authorize, inside the file and beyond it.
Specs was introduced Oct. 15 as a context container on Figma's canvas and remains in open beta. NextNow’s earlier analysis examined how that shared context gets reviewed. This week’s extension raises a narrower issue: access to an explanation, permission to alter a design and authority to change another system are different things. This analysis uses Figma’s documentation, not independent product testing.
Inquiry is not editing authority
Figma’s current agent documentation makes the first boundary explicit. A Full seat combined with file-editing access permits a user to chat and request changes. Users with View, Dev or Collab seats, or view-only file access, can chat without editing that file. Dev Mode has its own eligibility limits: it is unavailable on Starter plans and to View or Collab seats. Figma separately documents a Draft-files exception: anyone can chat with the agent and request edits in their drafts.
The guide also distinguishes preparing assets from exporting them. The agent can generate download links for a person to save assets, but the supported-features table does not describe the agent exporting those assets on the user’s behalf. An answer, a suggested implementation and a completed action should therefore not be treated as the same output.
Analysis: A product team could test this distinction with a developer reviewing two approved screen versions. Ask what changed and which component is relevant, then separately request a canvas edit. Record the permitted and refused actions rather than treating a useful explanation as evidence of broader authority. That exercise evaluates the account and task together; it is not a claim that asking a question should automatically unlock editing.
A connector crosses another boundary
The connector documentation describes a second interface. When Figma’s agent connects to an outside service, Figma acts as a client calling that service’s tools. By contrast, the Figma MCP server makes Figma file data available to other clients. Each person authenticates their own connectors, and access may depend on their account, plan, seat or role in the external service.
The same guide describes importing a Notion requirements document and, with write tools enabled, sending content back. It says the agent identifies a planned connector and tool and asks for consent, with individual tool permissions controlled in settings. Those settings matter: the guide also describes an Always run option, so a consent prompt should not be assumed for every operation.
Analysis: A prototype brief can be useful input without making every connected record an appropriate destination for changes. Design leads can identify the systems involved, the account used and the intended direction of each transfer before evaluating a connector-assisted task. Reading research, proposing an acceptance criterion and updating a project record deserve separate review questions. This is a proposed control scenario, not a report that a connector bypasses existing permissions.
AI access and training are separate
Figma’s administration guide separates access to AI features from content-training preferences and paid-credit access. Turning off content training does not disable AI capabilities. On Enterprise plans, organization administrators can set AI access at organization or workspace level; a workspace-specific setting applies to files there. A user’s seat and file access still constrain which features they can use. Paid-credit access is managed per person and is distinct from enabling AI features for a workspace.
The controls are not identical across plans. The guide says Professional and Organization accounts that previously disabled AI can re-enable it, but cannot subsequently turn that access off through the same setting. Enterprise administrators retain enable-and-disable controls. Those are documented configuration differences, not a guarantee about every piece of information moving through an AI-assisted workflow.
Figma’s Nov. 20 workflow article supplies a useful, limited illustration. Its fictional Pickup marketplace combines requirements, annotated frames and design-system guidance before generating an implementation plan. An engineer reviews and corrects the plan, then asks a designer for another change. The account says the team can review and test the implementation before rollout. It is a vendor-authored demonstration, not evidence of customer outcomes.
Analysis: The evaluation should keep three records distinct: what the design says, what the account can change and what the organization permits. That makes this release a reason to examine specific permissions, not to repeat an abstract promise that more shared context will produce a better product.
