Agentic AI
What happened
Figma launched Specs, collaborative documents that combine written instructions with designs, diagrams, and code on its canvas. Figma’s agent can work through comments, while coding agents can access specifications and their linked context through Figma’s Model Context Protocol server.
Why it matters
A capable agent can still build the wrong thing when decisions are scattered across documents and conversations. Specs makes the team’s intent a shared, editable input to execution, giving humans and agents a common reference as the work changes.
What’s next
Specs is rolling out in open beta to paid plans across web, desktop, and mobile. The practical test is whether keeping specifications beside the work reduces the corrections needed between design and implementation.
🛡️ BlueRock Screens the Agent Supply Chain
What happened
BlueRock launched its Agentic Trust Registry to analyze agents, skills, tools, and MCP servers before deployment. The company says developers can integrate its component checks into software build pipelines and use findings alongside its runtime monitoring and controls.
Why it matters
Every tool an agent gains adds another dependency to trust. The implication for buyers: evaluating the model alone leaves unanswered questions about the components that give it access and turn its decisions into actions.
What’s next
BlueRock is offering the registry for component evaluation and ongoing development checks. Its claimed detection advantage still needs independent validation; the useful test is whether findings identify exploitable problems without overwhelming developers with false alarms.
Generative & Enterprise AI
🔎 Anthropic Opens Its Doors to METR
What happened
Anthropic said it has onboarded six METR evaluators with hardware, office space, and internal access roughly comparable to its own risk-assessment teams, subject to defined exclusions. METR can publish assessments without an Anthropic veto, although Anthropic can require narrowly specified redactions that METR may publicly challenge.
Why it matters
Independent scrutiny becomes more meaningful when evaluators can inspect internal practices and publish uncomfortable findings. Access and publication rights are the consequential details here: they determine how much outsiders can learn and how much the public can hear.
What’s next
Anthropic is developing three evaluator-access tiers and hopes to embed more organizations at METR’s level. SecureBio currently has narrower access and publication rights; work with Accenture and Faculty covers evaluation, red-teaming, alignment, and safeguards.
Physical AI
🔋 Agility Targets a Twenty-Hour Robot Workday
What happened
Agility Robotics selected Nyobolt as Digit 5’s battery technology provider. Agility says the planned system will deliver 90 minutes of runtime from a nine-minute charge, supporting more than 20 hours of work per day; these are company projections, not independently verified fleet results.
Why it matters
A robot’s economics depend on the hours it spends working. Faster charging could improve utilization across warehouse and manufacturing shifts, provided the battery and the robot sustain that performance under actual workloads.
What’s next
Early access to Digit 5 with Nyobolt batteries is expected in the first half of 2027, with availability increasing over the following six to nine months. Customer deployments will test whether the projected charging gains translate into productive uptime.
💡 Bottom Line
Useful autonomy depends on what surrounds the model: shared instructions, trustworthy components, independent scrutiny, and reliable hardware. Today’s announcements strengthen different parts of that operating system. The commercial advantage will come from proving those pieces work together under everyday conditions.
⚙️ Try It Yourself
Give the agent one source of truth. Then see what changes.
Pick a small project you are working on — a landing page, a workflow, a dashboard, or a product update.
Write a short spec with four things:
Goal: what should be built
Context: the key design, notes, or constraints
Tools: what systems or components it can use
Boundaries: what should be reviewed by a human
Then give that spec to ChatGPT, Claude, or a coding agent and ask:
******
“What would you build, what would you need to check, and where could this go wrong?”
******
If you want to go one step further, ask it to list the outside tools or dependencies it would rely on and then review that list as if you were screening an agent supply chain.
Insight: Useful autonomy depends less on one smart model than on shared instructions, trusted components, and systems that can actually hold up in the real world.