Figma details agent guardrails with specs, libraries and access

At Config India 2026, Figma presented a more explicit framework for how its own agent and third-party agents can work inside its platform. Across its release notes and a report by TechCrunch, the company described a set of tools meant to give agents more context and more constraints: a new Markdown-based editor called Specs, guidelines attached to design libraries, code sources, centralized design systems, and wider support through Figma MCP. The common thread is governance. Figma is not only adding AI features; it is also trying to define how teams can shape agent behavior with rules, files, and shared references.

The emphasis on guardrails appears tied to a limit Figma itself acknowledges. TechCrunch reports that, after making its agent on canvas available to a wider user base in the past week, Figma is also providing tools to steer it toward more predictable outcomes because AI is non-deterministic. That framing matters. Rather than presenting agents as fully reliable by default, the company is positioning context and system rules as a way to reduce variability.

A native place for intent and context

The central announcement is Figma Specs. In Figma’s release notes, Specs is described as a native home for context that brings written intent, visual exploration, and technical details together for teams and agents. TechCrunch adds that this is a Markdown editor that lets users define their design language in a simple way. According to that report, Specs can exist as a standalone file in Figma or alongside the canvas, and users can reference other Figma elements such as canvas elements, FigJam diagrams, code layers, and designs.

Taken together, the two sources suggest that Specs is designed to do more than document a project for people. It is also meant to become machine-readable context for AI systems working in Figma. The release notes do not go into the same technical detail as TechCrunch on embedding or referencing other Figma objects, but both sources align on the broader role: agents should be able to consult the same written and visual context as human collaborators.

Availability is clearer here than for some other announcements. Figma says Specs is rolling out now in open beta on paid plans. That distinction is important for governance discussions inside organizations: the feature is not described as generally available, and it is not listed for all plans.

Libraries, guidelines and code as agent constraints

Beyond Specs, Figma is also formalizing other layers of control. Its release notes say teams can add guidelines to published design libraries so the agent follows established patterns and avoids anti-patterns. This feature is rolling out now in open beta. Figma also says code from GitHub repositories or npm packages can be imported as code sources to help the agent build and edit components and code layers that reflect real design systems. That feature is rolling out now in closed beta.

TechCrunch presents a similar picture, saying teams can bring in more context about design, including design libraries, code libraries, and guidelines to manage context for the design system, and that agents can refer to this centralized system to adhere to rules. Figma’s own release notes are more specific on product packaging. They describe centralized design systems as bundles of design libraries, code libraries, and guidelines grouped into a single design system so maintenance is easier for authors and adherence is easier for everyone. But Figma labels centralized design systems as preview only.

• Specs: open beta on paid plans.
• Guidelines in published design libraries: open beta.
• Code sources from GitHub repositories or npm packages: closed beta.
• Centralized design systems: preview only.
• Figma MCP with 20 new partners: available now.

This breakdown matters because the governance story is spread across features at different maturity levels. Some controls are already broadly usable for paying customers, some are limited to selected users, and some remain only a preview. Any company assessing how much policy and process it can actually enforce with Figma agents today will need to separate announced direction from present-day availability.

Code layers widen the agent’s operating range

Figma is not only tightening instructions; it is also expanding what agents can do. Code layers, first announced in private beta in June according to TechCrunch, are now rolling out in open beta, according to both sources. Figma describes them as a way to create interactive, code-backed prototypes on the canvas by converting a design frame, importing a GitHub repository, capturing a website with the Figma Chrome extension, or prompting from scratch.

TechCrunch adds another governance-relevant detail: users can ask the agent to turn a design into code or change part of the code to iterate on a prototype. In other words, the agent is moving closer to a build surface, not just a brainstorming assistant. Figma’s release notes reinforce that direction by saying Figma MCP can now be used from external AI tools to create code layers, edit specs, or generate images on the canvas.

That combination has two implications. First, context controls become more significant when the agent can affect code-backed prototypes rather than only static design artifacts. Second, interoperability becomes a governance issue of its own. Figma says MCP now has faster performance, support for more Figma products, and support for 20 new partners. TechCrunch names Open Code, Dia, Grok Bot, and Sarvam among the added tools. Figma’s notes do not list those names in the excerpt provided, so that partner detail comes from TechCrunch.

Payload integration and the shift from prototype to app

Another part of the announcement extends beyond design governance into application delivery. Figma says Make now natively scaffolds a Payload backend, providing a CMS, database, media, APIs, authentication, and an admin panel, with Figma hosting it so content and data live in the same place as the designs that shape them. This is rolling out across all plans, according to the release notes.

TechCrunch similarly reports that Figma is integrating Payload with Figma Make for generating dynamic websites and directly managing content. The article also notes that Figma acquired the open-source headless CMS Payload last year. While this integration is not framed primarily as an agent-governance feature, it strengthens the broader product logic behind the announcements: design context, code context, and content context are being pulled into one environment where agents can act.

What is confirmed, and what remains directional

Across the sources, the core message is consistent: Figma wants teams to define how agents behave by giving them explicit context and rules. Still, caution is warranted on two points. First, several capabilities are not fully available. Open beta, closed beta, and preview only do not mean the same thing operationally. Second, some implementation details appear only in one source. For example, TechCrunch describes Specs as Markdown-based and says it can reference embedded Figma elements, while Figma’s release notes focus on the broader product purpose rather than those exact mechanics.

The sources also differ in emphasis. TechCrunch highlights the rationale that AI is non-deterministic and that Figma is trying to steer outcomes toward predictability. Figma’s release notes emphasize workflow and product coverage, describing ways to build and collaborate with code, capture and share context, and build with real data and content. These are compatible accounts, but they are not identical.

For English-speaking companies, the practical change is that Figma is giving teams more concrete levers to decide what an agent should see, which rules it should follow, and where it is allowed to act. Today, that can mean documenting intent in Specs, attaching guidance to published libraries, and testing code-backed workflows in open beta, while recognizing that some controls remain in closed beta or preview. In short, organizations can start treating agent behavior less as an ad hoc prompt exercise and more as a governed part of the design system, but they will need to map that ambition against the current rollout status of each feature.