Field report · 21 June 2026
The agent moved into your tools, and started showing its work
This week the agent stopped handing back a mockup and took up residence inside the tools. Cursor's automations screenshot what they changed, Lovable became a connector you drive from Claude, and a new Google paper argues the craft is the harness, not the model.
Agentic tools spent the last year learning to generate. This week they were learning to stay: to run inside the project you already have, act on it directly, and leave evidence of what they touched. Read together, the three describe a week where the agent stopped being a window you open and became a hand inside the workshop.
Art Nouveau Editorial
The briefing as an illustrated Art Nouveau feature: a decorated margin, a gold drop-cap, and whiplash ornament framing one calm reading column on deep aubergine.
Art Nouveau built its pages from the margin inward. Ornament grew along the edges, framed the text, and left the reading column calm. Translated to the web it becomes an editorial feature: a decorated left rail carries the nameplate, the edition meta, and a contents index, while the story runs as a single column with a gold drop-cap and numbered chapter heads. The illustration sits inside a gold keyline the way a Mucha panel sits in its border, and the color stays in the ornament so the words keep their contrast.
Cursor's automations learned to screenshot their own work
Cursor 3.8 gave its Automations, the always-on agents that run on a schedule or a trigger, two changes worth setting up for. A new /automate skill lets you describe the job in plain language inside an ordinary agent session and have Cursor assemble the triggers, instructions, and tools for you. The larger one is computer use: a cloud automation can now open a browser inside its sandbox, click through your staging URL, and capture a screenshot or a short recording of what it did, returned next to the pull request it opened.
The effect is small and important. An agent that simply reports a green check asks for trust; an agent that hands back a picture of the working page has spent that trust down to something you can see. The change ships on by default for new automations, alongside fresh GitHub and Slack triggers.
cursor.comAn agent is a model plus a harness, and the harness is your job
Addy Osmani pulled the load-bearing ideas out of a Google whitepaper he co-wrote on how agents reshape the software lifecycle, and the framing that travels is plain: an agent is a model plus a harness, split roughly ten percent model and ninety percent harness. The harness is everything wrapped around the model, the rule files, the tools, the sandboxes, the orchestration that routes between sub-agents, the checks that notice when it drifts. So when an agent does something dumb, you debug the harness first, because most agent failures turn out to be configuration you can fix today rather than a model you have to wait on.
The same essay draws the cleanest line yet between vibe coding and engineering, and it is verification. Tests cover the deterministic parts, evals cover the rest, and the rule he leaves you with is to set the bar at the eval, not the demo. A demo proves an agent can work once; an eval proves it works on purpose.
addyosmani.comLovable turned itself into something you drive from Claude
Lovable opened its MCP server to Claude on every plan, including the free one, which changes what the builder is: less a place you visit, more a connector other agents operate. Connect it once and your Lovable workspace shows up inside Claude, where you can ask it to audit your projects, query live data from an app you already shipped, or kick off a fresh build without leaving the chat. From Claude Code the same server exposes /build, /deploy, and /db as single terminal commands, and a prototype handed over from Claude Design lands in Lovable with auth, a database, and a live URL already attached.
docs.lovable.devThe agent stopped describing the work and started doing it where the work lives.
Prompt Lab
Recreate today's art directionDesign a single web page as an illustrated Art Nouveau editorial feature presenting a short briefing of linked stories. Page archetype: a long-form magazine feature with a decorated margin, not a dashboard, blog, or card grid. Ground: deep aubergine jewel (#271629) with a lifted plum surface (#32203A). Text is warm parchment (#F2E7D2) at 20px, line-height 1.72, measure near 64 characters, never justified. Use a Mucha-register palette: antique gold (#C8A24C) for links, rules, drop cap, and all ornament; olive-sage (#9FA977) and muted mauve (#C2A6BD) to color-zone the chapters. Keep every text color WCAG AA on aubergine; let color live in the ornament, not under the words. Layout: an asymmetric two-part page. A decorated left rail (sticky on desktop, stacked on mobile) holds an ornamental nameplate, edition metadata as label-value rows, and a contents index. The story runs as one framed reading column: a thin gold keyline border with whiplash corner flourishes, a Yeseva One headline, a deck, a single illustrated plate, then a lead paragraph opened by one gold drop-cap. Sections are numbered chapter heads (I, II, III) with a spaced-caps label and a whiplash hairline divider in the chapter's zone color; each story is a kicker, a linked headline, one or two plain paragraphs, and the source domain. Exactly one italic display pull-quote. Type: Yeseva One for the display headline, chapter numerals, and drop cap; Lora for body; Cormorant Garamond italic for the one pull-quote and the art-direction summary; IBM Plex Mono for labels, metadata, and the prompt block. Include one generated plate: a symmetrical Art Nouveau panel with a gold sunburst halo, a stylized iris, mirrored whiplash scrollwork, sage leaves, and small gold-keyline cartouches of abstract page layouts, all flat color fields with gold keylines and no readable text inside the image. Add one restrained micro-interaction, a link underline that warms to gold on hover, behind a prefers-reduced-motion guard. No gradients, no glow, no neon, no pill-shaping everything, no accent left-border stripes, no fake search fields or toggles, no emoji.
Works in: Beaver Builder AI, v0, Lovable, Cursor, Figma Make.
Three releases, one direction: the agent left the chat window and set up inside the tools themselves, where it can act on the real project and show what it changed. The craft that remains, the paper argues, is the harness you build around the model and the checks you hold it to.
Sources
- Improvements to Cursor Automationscursor.com
- The New Software Lifecycleaddyosmani.com
- Build in Lovable from Claudedocs.lovable.dev
- Lovable MCP serverdocs.lovable.dev