Appiness #21

Tech

• A study reported by Ars Technica on 10 October finds that coding agents write more code without teams shipping more software. The gain disappears in human review before the code can be released.
• OpenAI has disclosed four incidents from testing and internal use in which agents did things they were not permitted to do. In one, an agent got into an internal system, changed it, kept its access and downloaded 147,451 rows from non-public tables.

Design

• Figma launched Specs at its first Config in India on 14 October. Specs is a Markdown editor where a team describes its design language, and both designers and agents read it. Alongside it came a shared home for design libraries, code libraries and guidelines that the agent follows when it builds.
• Code layers, which put code inside a design for interactive prototypes, moved from private to open beta. Figma's MCP server added support for more than 20 new tools, and Payload, which Figma acquired last year, is being built into Figma Make.
• Figma's chief design officer Loredana Crisan said 17 times more lines of code are being written today than before, and that much of that code is never meant to ship. It is written to find out what the product should be.

Business

• Oslo-based Solvera has raised $6.3 million in pre-seed funding, led by Seedcamp. Investors include Obos Ventures, Momentum and PSV Tech, with Spacemaker co-founders Anders Kvåle and Håvard Haukeland as angels. Solvera is the engineering firm itself: agents handle a growing share of mechanical, electrical and plumbing design for buildings, and engineers make the key judgement calls. It already works with the contractor Betonmast and plans to open an office in London.
• Oslo-based Neonomics has raised $4 million from existing shareholders. The round follows a restructuring that cut its cost base by almost 50 percent, and new customer acquisition is up more than 60 percent since the start of 2026.
• Accenture will fund about $29 per employee per month of Claude usage from a central budget. Spending above that moves to project or team budgets. The same week, figures showed junior hires falling from 27 to 9 percent of Anthropic's technical hiring, even as technical hiring grew overall.

The Apps take

This week's stories point the same way. Writing code has become cheap. What costs money is deciding what to build and standing behind what was built.

The Ars study finds the gain again in review. Figma answers with Specs, because an agent only builds the right thing when the rules are written down. OpenAI's incidents show what happens when nobody approves an action before it is carried out. Solvera builds a whole company around the same split: the agent designs, the engineer owns the judgement.

If a coding agent triples the number of changes, you have tripled the queue at whoever approves them. That is where the risk sits when software acts on its own.

Before you turn up the speed on your next agent, write down three things: what it should follow, who approves its work, and how large a change can get before it must be split. Without those three answers, a faster agent only gives you a longer queue.