Singapore tests agent governance as Cohere and Figma expand controlled rollouts

The May 20 AI Wire brief focuses on agent systems moving toward operational use under tighter governance, deployment and availability constraints. Singapore agencies and Google reported findings from a public-sector AI Agents Sandbox, and IMDA updated its agentic-AI governance framework. Cohere released Command A+ as an Apache 2.0 open model for enterprise and private deployment. Figma began rolling out a native design agent in closed beta.

The common thread is controlled adoption. The sources support new activity in government sandboxes, enterprise model deployment and design-workflow agents. They do not support independent claims that the sandbox outcomes are validated at scale, that vendor benchmarks will reproduce in production, that beta design-agent access is broad, or that education deployments have measured learning outcomes.

Top developments

Singapore and Google publish AI Agents Sandbox findings

Singapore's Cyber Security Agency said on 2026-05-20 that an AI Agents Sandbox run with Google, GovTech Singapore and IMDA examined computer-use agents in public-sector settings. The sandbox focused on automated quality assurance for government digital services, AI safety testing and social-assistance application workflows.

CSA said the sandbox showed potential for agent-assisted quality assurance, larger-scale AI safety testing and help navigating complex public-service applications. The same release identified risks around human oversight, cybersecurity, privacy, data protection, customisation and control. It also flagged indirect prompt injection as a risk that could make an agent perform unintended actions, including remote code execution.

IMDA separately said it updated the Model AI Governance Framework for Agentic AI on the same date. The update adds case studies and best practices after feedback from more than 60 organisations, including guidance on multi-agent systems, third-party agents and automation bias. IMDA frames the framework as guidance, not binding regulation, and says humans remain ultimately accountable.

Operational read: Singapore is using sandbox findings and governance guidance to define how agents should be tested before broader deployment. The brief should not treat participant-reported sandbox outcomes as independent validation or convert source-identified risk themes into confirmed incidents.

Cohere releases Command A+ as an Apache 2.0 enterprise model

Cohere announced Command A+ on 2026-05-20 as an open-source enterprise model available under Apache 2.0. Cohere describes it as a sparse mixture-of-experts model with 218 billion total parameters, 25 billion active parameters, 128K input context, 64K maximum generation, image input, tool use, reasoning and support for 48 languages.

The source-backed change is Cohere's move to combine multimodal, multilingual, reasoning and agentic capabilities in one Command-family model for enterprise use. Cohere and its model card also describe W4A4 quantized deployment guidance, including a minimum configuration of one B200 or two H100 GPUs for that variant.

The performance and efficiency claims need attribution. Cohere reports benchmark gains over prior Command models, improved internal North evaluations and near-lossless or negligible-quality-difference quantization claims. Those claims are vendor-reported unless independently reproduced. Actual serving results will depend on workload, stack, concurrency, quality thresholds and hardware.

Operational read: Command A+ matters because enterprise model releases are increasingly framed around local or private deployment, tool use, long context and lower hardware footprints. Apache 2.0 licensing supports commercial reuse, but it does not by itself prove training-data transparency, regulatory compliance, customer outcomes or sovereign-AI readiness.

Figma starts rolling out a native design agent in closed beta

Figma announced on 2026-05-20 that it is rolling out a design agent built into Figma Design. The agent appears on the canvas and in the left rail, can work with selected layers and can use design-system context such as components, tokens, variables, styles and libraries.

The practical delta is that Figma is moving agentic work into its primary design surface. Figma says the agent can generate and remix designs, automate bulk edits, get design feedback, update content, use libraries and answer Figma Design questions. The help documentation says users can undo agent changes and that chat history is private within the file.

Availability is limited. Figma says the agent is in beta, began rolling out on May 20 and is gradually enabled at the team or organisation level. It is available to Professional, Organization and Enterprise plans for eligible seats, with Dev or Collab seats limited to Drafts. The help page says Starter, Government and Education plans are not eligible, waitlist signup does not guarantee access and Figma does not provide exact access timelines. The beta is free and does not consume AI credits, but Figma says standard AI credit usage will apply after general availability.

Operational read: design agents are becoming embedded workflow features rather than only external experiments. Teams should separate supported beta capabilities from claims about design quality, autonomy, cost and general availability.

Watchlist

OpenAI adds Singapore to Education for Countries

OpenAI said on 2026-05-20 that Singapore is joining its Education for Countries program and that it is seeking a new cohort of country partners. OpenAI describes the program as government-led and research-based, with pillars around research-driven deployment, localized AI tools for learning and teacher training. The same post says OpenAI will launch an educator engagement track called OpenAI Luminaries.

This remains a compact watchlist item because the public-sector education risk surface is high and the reviewed sources do not provide country-specific implementation contracts, procurement terms, privacy details or measured learning outcomes for Singapore. OpenAI's related May 19 Singapore announcement says it plans to work with Singapore's Ministry of Education and GovTech on AI-enabled learning tools and use cases. Singapore MOE's public AI-in-education page gives broader context on its responsible-use approach and ethics framework, but it does not confirm the specific May 20 Education for Countries details.

Operational read: education is becoming another government-partnership surface for frontier AI vendors. Coverage should keep vendor and government claims attributed, especially where minors, learning outcomes, privacy, equity, procurement and teacher labour are involved.

France / EU watch

No fresh France/EU-specific operational item met the source-backed threshold for this May 20 brief. OpenAI's Education for Countries update includes Greece and Estonia as program context, but the selected fresh item is broader public-sector education positioning rather than a new France/EU regulatory or deployment development.

Known caveats

• Singapore sandbox outcomes are source-reported and should not be framed as independently validated or generally proven across all public-sector services.
• IMDA's agentic-AI framework update is guidance and case-study material, not binding regulation.
• Security and privacy risks in the Singapore item are source-identified risk themes, not confirmed incidents.
• Cohere benchmark, efficiency, quantization, throughput, latency and sovereign/private-deployment claims are vendor claims unless independently reproduced.
• Apache 2.0 licensing does not establish training-data transparency, regulatory compliance or customer production outcomes.
• Figma's design agent is in closed beta with gradual team- or organisation-level rollout; waitlist interest does not guarantee access.
• Figma has not provided a general-availability date or fixed post-beta AI-credit cost for the design agent in the reviewed sources.
• OpenAI's Education for Countries update lacks reviewed detail on Singapore-specific implementation contracts, procurement, privacy safeguards or measured learning outcomes.
• No France/EU-specific operational delta is asserted for this edition.

Sources

• Cyber Security Agency of Singapore, AI Agents: Insights from the Singapore Government and Google Sandbox: https://www.csa.gov.sg/news-events/press-releases/ai-agents--insights-from-the-singapore-government-and-google-sandbox-/
• IMDA, Updated Model AI Governance Framework for Agentic AI: https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2026/updated-model-ai-governance-framework-for-agentic-ai
• Google, Google extends partnership with Singapore government on AI: https://blog.google/company-news/inside-google/around-the-globe/google-asia/singapore-government-partnership/
• Cohere, Introducing Command A+: https://cohere.com/blog/command-a-plus
• CohereLabs Command A+ model card: https://huggingface.co/CohereLabs/command-a-plus-05-2026-w4a4
• Cohere model documentation: https://docs.cohere.com/docs/models
• Figma, The Figma design agent is here: https://www.figma.com/blog/the-figma-agent-is-here/
• Figma Help Center, How do I access the AI agent beta in Figma Design?: https://help.figma.com/hc/en-us/articles/34932042346775-How-do-I-access-the-AI-agent-beta-in-Figma-Design
• Figma Help Center, Work with the AI agent in Figma Design: https://help.figma.com/hc/en-us/articles/37998629035799-Work-with-the-AI-agent-in-Figma-Design
• OpenAI, The next phase of OpenAI's Education for Countries: https://openai.com/index/the-next-phase-of-education-for-countries/
• OpenAI, Introducing OpenAI for Singapore: https://openai.com/index/introducing-openai-for-singapore/
• Singapore Ministry of Education, Artificial intelligence in education: https://www.moe.gov.sg/education-in-sg/educational-technology-journey/edtech-masterplan/artificial-intelligence-in-education