Laura: Okay, Vercel dropped AI SDK 7 this week and it's actually solid infrastructure work, not hype.
Harper: Yeah, I saw the release notes. Sixteen million weekly downloads, so this is the layer a lot of people are building on.
Laura: Right. And what's interesting is they're not adding new reasoning or a novel agent loop—they're shipping what teams are actually hitting production walls on. Approvals, durability, tool context scoping, file upload deduplication.
Harper: Mm-hm.
Laura: The reasoning thing is slick though. Every frontier model supports configurable reasoning effort now—OpenAI, Anthropic, others—but they all expose it differently. SDK 7 standardizes it to a single `reasoning` option that maps to the provider's native setting. One line instead of digging into provider-specific options.
Harper: That's nice but that's mostly API wrapping. What's the actual problem it solves?
Laura: Developer experience. You don't have to learn five different reasoning APIs; you set `reasoning: 'high'` and it works across providers.
Harper: Fair. What about the tool stuff?
Laura: Tool context is the interesting piece. Tools are increasingly built independently—third-party companies ship tools that agents use. Problem is, those tools need secrets and config that shouldn't leak to other tools or the model. AI SDK 7 adds a context schema per tool, so you can say 'this weather tool needs an API key, and ONLY this tool gets access to it.' The context is scoped. And then there's file uploads. If you're doing multi-step inference—like a coding agent that ref
Harper: That's real. Stateless inference sends a lot of dead weight if you're not deduplicating. Okay so file and skill uploads are shipping real bandwidth wins. What about durability?
Laura: This is the one I think teams actually need. New thing called WorkflowAgent. Agent runs span multiple steps, sometimes they wait for human approval—if the server restarts or you deploy in the middle of that, you've lost state and you start over. WorkflowAgent makes agent execution durable across process restarts, deployments, interruptions. State survives. It's a workflow layer underneath the agent loop. Tool calls, approvals, callbacks, runtime context—all serializable acros
Harper: Okay, that's actually useful.
Laura: Right? And then there's tool approvals. You can gate tool execution with simple user approval, or a function that auto-approves, auto-denies, or escalates to the user. You can do it per tool or catch-all. They even added HMAC-signed approvals so you can't forge them.
Harper: That's the control surface people have been asking for. You ship an agent, it tries to do something risky, you need to be able to say no.
Laura: Exactly. This is infrastructure that actually maps to what shipping teams care about.
Harper: Mm-hm… but here's the thing. Tool approvals and durability are real, but they're solving a different problem than the one I think teams are actually stuck on.
Laura: What do you mean?
Harper: The hard part—the one that's actually killing teams—is specifying what success looks like. What should the agent NOT do? When should it stop? How do you verify it worked? No SDK feature ships that for you. That's constraint expressibility, and it's still on the team.
Laura: Okay, but that's not AI SDK's job. That's harness design, which we've talked about before—it's on the team building the agent. The SDK is the layer underneath. It's saying 'here's how you gate execution, here's how you keep state, here's how you talk to any model.' The constraints are your problem.
Harper: Right, I'm not saying it's broken. I'm saying it's good infrastructure but it doesn't move the needle on the actual bottleneck. It's necessary but not sufficient.
Laura: Fair. What about MCP Apps? That one's a bit different.
Harper: MCP Apps is clever. MCP servers can now separate model-visible tools from app-only tools, and render a UI in a sandboxed iframe. So the agent sees the tools it needs, the user sees a dashboard or control panel.
Laura: That's smart because not everything should be model-visible. Some things are just for humans to configure or review.
Harper: Yeah. But this assumes teams are already building with MCP. If you're not, this is a nice-to-have.
Laura: True. But MCP is becoming the standard for agent tooling, so it's forward-looking.
Harper: Fair.
Harper: Yeah, no, I agree. It's solid infrastructure. I just want to make sure people don't confuse 'shipping good tooling' with 'solving the agent bottleneck.' They're different things.
Laura: Totally fair. This is the layer that makes it possible to build a good harness. It's not the harness itself.