Vibe Code Brief, October 18: AI Agents Learn To Take Direction Mid-Task

Sunday edition, four items, one through-line: AI tools are racing to make human judgment easier to apply. Agents you can steer mid-task, design rules written down where agents can read them, a funded workspace for team-driven agents, and now even CAD has an MCP door on it.

1. Codex CLI 0.163: you can now steer the agent mid-response

Image: OpenAI, via Codex CLI documentation

What happened: OpenAI shipped Codex CLI 0.163 and made steering the default. Per the release notes, you can send a follow-up while Codex is still responding and it corrects course immediately, and interrupting a turn also stops the active Code Mode cells and their nested tool calls. The release also adds searchable settings and configurable shortcuts.

Why it matters to designers and builders: the biggest tax in agent work is waiting for a full wrong answer before you can fix it. Mid-flight correction is the same tight feedback loop you already run in critique: small corrections, early, cheap. The tools are converging on "interruptible beats impressive."

What to do: update, then rebuild your habit: when the agent drifts, redirect in one line instead of letting the full run finish. Your iteration speed roughly doubles.

2. Figma at Config India: Specs for humans and agents, code layers open beta

Image: Figma, via Introducing Specs: Markdown as a Medium

What happened: at its first Config India, Figma announced Figma Specs, a markdown editor for defining your design language: rules, libraries, guidelines, with references to canvas elements and code, sitting on the canvas as shared context. Code layers, which put code-backed interactive prototypes directly on the canvas, moved from private beta to open beta. Figma also integrated Payload CMS with Figma Make and extended MCP support to 20+ tools (TechCrunch). Chief design officer Loredana Crisan's framing stuck with me: 17x more lines of code are being produced, but only 1.3x more releases. Most code is not built to ship, it is built to discover what the product is. It is built to design.

Why it matters: written design rules are now machine-readable context that every agent can be held to. The teams with excellent, documented design language get proportionally more out of AI than the teams whose taste lives in one director's head.

What to do: write your design rules down this week as a markdown file, even a rough one, and point every agent (and new hire) at it. Budget: an afternoon. Returns: every future generate.

3. LightSprint raised $6.1M to make coding agents a team sport

Image: LightSprint, via lightsprint.ai

What happened: Y Combinator startup LightSprint raised $6.1 million to build a shared workspace where product managers, designers, and engineers run Claude Code, Codex, and Cursor agents together: shared context, shared skills (your design.md, your review.md), sandboxed previews you can click through, and approvals before anything ships.

Why it matters: today most agent sessions live in one developer's IDE and the rest of the team finds out later. A shared surface means designers can propose and click through changes without a handoff queue, on the real codebase, with guardrails.

What to do: if your agents are trapped in one person's terminal, trial a shared workspace model with your next small feature, one designer, one dev, one agent, and see where the friction actually goes.

4. SolidCraft: open source parametric CAD with an MCP door

Image: SolidCraft, via the project's GitHub repository

What happened: Hackaday flagged SolidCraft, an open source parametric 3D CAD application written in pure Rust, running on desktop and in the browser, with automation exposed through MCP (GitHub).

Why it matters: the vibe-coding wave keeps eating heavier software. Parametric CAD was one of the last "no way an agent touches that" categories, and it just grew an agent door. If you design physical products or 3D assets, your automation surface just expanded. And if you are a UI designer, note the pattern: every serious tool is becoming agent-accessible, which means every serious tool will need designers who can spec, steer, and judge.

What to do: watch it, and try it now if anything you ship has three dimensions.

Our take

Four different products, one message: the interface for AI in 2026 is not a better chat window, it is steering and shared context. Codex optimized for "correct me early." Figma is betting your design language becomes a document. LightSprint assumes agents work best as a team surface. SolidCraft proves even CAD wants to be automated. None of this makes judgment less valuable. All of it makes judgment faster to apply. Write your rules down, keep your reps, steer hard.

Building something from today's news? Base44 is still the fastest way I know to take a design habit to shipped product. Want reps? Daily Vibe Code.

Related on NewForm

• United's CEO Vibe-Coded a Betting App Between Gym Sets. Here's What That Means for Designers
• A Vibe-Coded Site Got Hacked 6 Days After Launch. It Was Not the AI's Fault
• 10 Trustworthy Fintech Apps Designs to Learn From (October 2026)