Skip to content

Appiness

Appiness #20

Anthropic put an agent on watch over its coding agent, Figma started charging for its own, and Norway's 2027 budget sets aside NOK 30 million to evaluate AI models.

By Apps AS3 min read
Two orange triangles meeting at their points like an hourglass on a pale cream background

Tech

  • On 1 October Anthropic added "You should know" to Claude Code 2.1.287: a side agent that watches the session and flags what you or the main agent might have missed. The same week, /code-review got a setting for how many findings it reports.
  • Google's product security team has described PageBreak, an internal security agent that has found more than 500 XSS vulnerabilities in Google's own web applications in 2026. The agent confirms each finding by running real payloads against the application before reporting it, so nobody has to sift false positives by hand.

Design

  • The Figma agent left beta on 6 October, together with the Weave tools. Whoever owns a design library can now upload guidelines as markdown files, and the agent follows them when it builds with the components.
  • From the same day, the agent draws AI credits from the same monthly pool as Figma Make. Figma's own examples are around 30 credits to swap fonts and more than 300 for a five-screen mobile checkout flow. A full Professional seat gets 3,000 credits a month.
  • Organization and Enterprise plans can now connect to several GitHub organizations at once.

Business

  • On 7 October the Norwegian government presented its 2027 budget proposal, with NOK 121.3 million for six AI initiatives and about NOK 2.2 billion for new digitalisation initiatives, up from NOK 1.6 billion this year. The list includes NOK 25 million for a shared generative AI platform for the public sector, NOK 20 million for agentic AI in licence processing at the energy regulator NVE, and NOK 30 million for the Norwegian Digitalisation Agency to evaluate AI models and work on AI safety.
  • Denmark's PSV Tech has closed EUR 56 million for its second fund, which backs companies in Denmark, Sweden and Norway in AI, software and digital infrastructure before they find product-market fit. Partner Helle Uth says the company with the best AI product is not necessarily the one that wins.

The Apps take

Every story this week is about control. Anthropic puts an agent on watch over the agent. Google makes its security agent prove each finding before a person sees it. Figma lets the library owner write the rules the agent follows, and charges per task. The Norwegian state sets aside NOK 30 million to evaluate the models it is about to use.

Generation now has a price you can read off a table: more than 300 credits for a checkout flow. Control is in no table. It costs a person who reads what the agent proposed and stands behind the result. That cost decides whether an agentic workflow pays off, and it is rarely in the calculation. It is also where the risk sits when software acts on its own.

Before you adopt the next agent, decide who approves what it does, how long the approval takes, and who does it when that person is on holiday. Without those three answers, you have moved the work from the person who writes to the person who reads.

Related articles