Vercel AI SDK 7, Figma Config 2026, and LangChain's Agent Development Lifecycle

Vercel — AI SDK 7#

Source: https://vercel.com/blog/ai-sdk-7 Credibility: High (first-party announcement, shipped today)

What happened: Vercel released AI SDK 7, a major version bump with breaking changes. The SDK is the most widely used TypeScript library for building AI-powered products — it abstracts over OpenAI, Anthropic, Google, and other providers and handles streaming, tool calls, and agent orchestration. A major version means real decisions were made about what to change and what to break. The changelog entry at https://vercel.com/changelog/ai-sdk-7 confirmed the release shipped today.

Key capabilities:

• Based on the title and prior SDK 6 trajectory: this is the TypeScript-first, provider-agnostic SDK that powers most Next.js + AI apps
• Prior versions added streamText, generateObject, tool calling, and multi-step agent patterns
• Version 7 follows the v0 and Eve agent framework work Vercel shipped in recent weeks — the SDK is increasingly the runtime layer for those higher-level tools
• GLM 5.2 Fast via Wafer was also added to Vercel's AI Gateway on June 24 (separate changelog entry), giving SDK users access to GLM 5.2 — the same model Lenny Rachitsky wrote about this week as a cost-effective Opus replacement

Why it matters for PMs: If your team is building any AI features on Next.js or TypeScript, there's a good chance you're using this SDK or will be soon. Major version bumps in foundational libraries create real migration decisions: do you stay on v6 or move to v7? The fact that Vercel is moving fast here (SDK 7 comes shortly after the Eve agent framework and Vercel Connect launches) signals they're positioning the whole stack — SDK, hosting, gateway, agents — as a unified platform rather than a set of loose tools. That's a meaningful strategic shift worth understanding before you make tooling decisions.

Critical questions:

• What are the actual breaking changes in v7? Migration cost matters more than feature additions for teams already in production.
• Does v7 improve or change the mental model for multi-step agents, or is it primarily performance and compatibility work?
• How does AI SDK 7 interact with the Eve agent framework Vercel shipped last week? Is Eve built on the SDK or alongside it?
• If you're using Anthropic or OpenAI directly (not via SDK), does v7 give you enough reason to introduce this abstraction layer?

Action you could take today: Pull up https://vercel.com/blog/ai-sdk-7 and check the migration guide. If your team has any TypeScript AI features, forward the breaking changes section to whoever owns that codebase before they get surprised.

Figma — Everything from Config 2026#

Source: https://www.figma.com/blog/introducing-figma-motion/ Credibility: High (first-party release notes, dated June 24)

What happened: Figma published a comprehensive Config 2026 roundup. The headline new product is Figma Motion — an animation and motion design tool built directly into Figma. Beyond that, the Config release included updates to Figma Make (the design agent), the Figma MCP server, and enterprise admin controls for AI credit usage. This is Figma's annual product moment, and the 2026 edition is clearly betting on motion design and AI-assisted design as the two growth vectors.

Key capabilities:

• Figma Motion: Native animation tooling inside Figma — this is the "one more thing" for Config 2026. No more exporting to After Effects or ProtoPie for motion work.
• Figma Make updates: Web search is now available in the Figma design agent (shipped June 18) — the agent can pull live web context and populate designs with real content
• MCP server additions: Figma Slides, uploaded fonts, and more are now accessible via the Figma MCP server (shipped June 16)
• AI credit usage API: Enterprise customers can now access AI credit usage data programmatically — critical for finance and IT teams managing AI cost at scale
• Workspace-level web publishing controls: Org admins can manage who can publish Make and Sites projects externally

Why it matters for PMs: Figma Motion is the kind of addition that changes who stays in Figma and who reaches for a specialized tool. If motion designers can stay in Figma for the full workflow, that's a meaningful stickiness play — and it expands the product's surface area into a category Figma didn't formally own before. For PMs building products with design-heavy teams, the practical question is whether your motion designers will actually adopt it or stick with their existing tools. The AI credit usage API is separately important: if you're at an enterprise with Figma, this is how you start managing and attributing AI costs before someone in finance asks you where the budget went.

Critical questions:

• How does Figma Motion compare to dedicated tools like After Effects, ProtoPie, or Jitter on actual fidelity? "Competitor-level" and "good enough for most teams" are different bars.
• The MCP server additions are expanding fast. Are there enterprise security implications to giving AI agents access to uploaded fonts and Slides content?
• Will Figma Motion be included in existing plans or gated as a paid add-on?
• The design agent now searches the web — how does Figma handle data privacy when the agent is pulling in external content during a design session?

Action you could take today: If you have a motion designer or prototyper on your team, send them the Config 2026 roundup and ask them to spend 30 minutes with Figma Motion. Their reaction will tell you more than any benchmark comparison.

LangChain — The Agent Development Lifecycle (+ Lyft Case Study)#

Source: https://www.langchain.com/blog/the-agent-development-lifecycle Credibility: High (first-party, multiple supporting posts published same day with specific production case study)

What happened: LangChain published a dense batch of agent engineering content today, anchored by a post called "The Agent Development Lifecycle" that formalizes a four-phase framework: Build, Test, Deploy, and Monitor. Alongside it: a Lyft case study showing they cut agent development time from months to weeks using LangGraph and LangSmith; a post on "The Art of Loop Engineering" explaining how to design agent harnesses for specific tasks; a "Give Your Agent Its Own Computer" post on sandboxed code execution; and an introduction to LangSmith Engine, which watches production traces, clusters failures into named issues, and proposes targeted fixes.

Key patterns (from the Agent Development Lifecycle framework):

• Build phase: Prototype with a model + tools, define the task boundary clearly before adding complexity
• Test phase: Evals aren't optional — LangChain now offers "Rubrics" (introduced today via separate post) for self-evaluation loops where agents check their own output against a rubric before returning results
• Deploy phase: Separate runtime concerns (sandboxing, compute, state) from logic concerns — the "give your agent its own computer" post argues that code execution agents need isolated environments, not shared infrastructure
• Monitor phase: LangSmith Engine closes the loop — it identifies failure clusters in production traces and proposes eval coverage for them (not just alerts, but proposed fixes)

From the Lyft case study:

• Lyft built a self-serve AI agent platform for customer support using LangGraph and LangSmith
• Development time went from "months to weeks" — the specific lever was LangSmith's observability making it faster to debug and iterate
• This is one of the cleaner enterprise adoption signals for agent frameworks: the value wasn't in the model, it was in the tooling around it

Why it matters for PMs: The agent development lifecycle framing is useful precisely because it's not new architecture — it's applying software engineering discipline (test, deploy, monitor) to a workflow that most teams are still treating as "prompt and pray." The Lyft case study is the real signal here: months to weeks is a velocity claim that's hard to fake, and the specific mechanism (observability accelerating iteration) maps directly to what PMs feel when they're waiting on engineering to debug a flaky agent. If you're planning any agent work in the next quarter, this four-phase framework is worth sharing with your tech lead before scoping begins.

Critical questions:

• "Months to weeks" is a strong claim. What was Lyft's specific baseline and what exactly changed? The case study headline is compelling but the detail matters.
• LangSmith Engine proposing "targeted fixes" from production traces sounds powerful. How much human review is needed before acting on those proposals?
• The Rubrics self-evaluation loop adds latency. For customer-facing agents, what's the latency budget tradeoff?
• LangChain is clearly positioning LangSmith as the production observability layer. How much of this is neutral tooling vs. a moat play to lock agent workflows into their stack?

Action you could take today: Read the Lyft case study (https://www.langchain.com/blog/lyft-built-a-self-serve-ai-agent-platform-for-customer-support-with-langgraph-and-langsmith) and pull the specific before/after metrics. If you have an agent project in flight, use the four-phase lifecycle as a checklist to identify which phase you're weakest in.