I’ve been reflecting lately on how fast the automation space has been evolving. Over the past few months, the lines have been blurring; between no-code and low-code, between AI-driven and traditional workflows, and between agentic and deterministic execution. Categories that used to sit neatly apart now bleed into one another.
Recent launches from two of my go-to platforms provide a great example of this shift.
Zapier SDK: 9,000 apps in your terminal
On 13 May, CEO Wade Foster opened the Zapier SDK to everyone. Install it and Claude Code, Cursor, Codex, or your coding agent of choice can call any of the 9,000+ apps in the Zapier catalogue and the 30,000+ actions embedded within them.
💡
What is an SDK? A software development kit (SDK) is a set of pre-built tools and code libraries that lets a developer (or a coding agent) interact with a platform without writing every line from scratch. Think of it as a ready-made interface: instead of handling authentication, API calls, and error handling yourself, the SDK wraps all of that into functions you can call directly. The Zapier SDK, for instance, gives your coding agent one clean way to reach 9,000+ apps without managing credentials or building integrations one by one.
What about Zapier MCP, which already let you connect your LLM to Zapier? MCP is for chatbots to interact with other apps in natural language, not code. The Zapier SDK is for builders (and their coding agents) writing scripts. Either way, Zapier acts as the integration layer, enabling AI to interact thousands of apps without generating and managing API keys for each one.
I rebuilt one of our internal workflows with it a few weeks ago. The original was a Sub-Zap that linked every incoming email or meeting in our Notion CRM to the right Contact records. It had 24 steps: paths, filters, a loop, an AI step to classify whether an address belonged to a real person or a generic service account, and more. It worked, but with all those steps, it was a beast to troubleshoot and iterate on. The refactored version is just one Code step plus the start and return from the Sub-Zap. It’s all the same inputs, outputs, and logic, but now that it’s expressed in code, debugging and editing it is lightning fast with Claude Code.
Next-Gen Zaps: the SDK becomes the product
Zapier is now testing something that goes a step further. In private beta (which I have started testing), you can use Cursor or Claude Code to build and edit Zap workflows directly; no visual editor required. The internal name is Durables, and it’s a complete rebuild of the core Zapier platform on top of the Zapier SDK as the foundation.
It works pretty much as you’d expect: you describe what you want in natural language and the agent picks the right app connections, wires the steps, handles branching and error logic, and deploys. There’s still a visual workflow diagram, reminiscent of the traditional Zap editor view, but all changes are made by editing the code rather than dragging and dropping or clicking into fields. You can click in and edit the code yourself, ask your external coding agent to do it, or use the built-in AI agent. The workflow diagram is now just a visual representation for our human brains, but the actual logic all runs as a single code block.
The future of building and editing Zaps
What makes this different from the SDK on its own is that Zapier hosts and executes the workflows once they’re built, rather than you and your coding agent just running them locally from your machine. It’s the SDK and the server, together.
It’s clear this was the vision Zapier had in mind when they built the SDK. Previously they had an AI co-pilot in the Zap editor, but it was super slow and inconsistent, because the original platform was designed for visual no-code building and not optimised for how LLMs work. Now, the entire product has been rebuilt from the ground up for the era of AI coding agents, and the vision of vibe automation is close to being realised.
Notion Workers: deploy code without leaving Notion
Around the same time, Notion announced its Developer Platform, headlined by Notion Workers: small programs you deploy through a new CLI, hosted and run on Notion's infrastructure.
Workers do three things that previously needed a separate automation tool:
Sync data from external services into Notion databases on a schedule (I wrote about this in depth here).
Give Custom Agents new tools with deterministic execution, at a fraction of the LLM token cost of having an agent reason its way through the same task.
Receive webhooks from anywhere, so any app can trigger an automation in Notion. The webhook does not have to come from an external service; it can come from within Notion itself, including from a legacy database automation. That lets you combine the two: the database automation catches the event, the Worker does the rest.
We have a dozen Workers running in our Notion workspace already. One fires whenever a new page is added to our meetings note database, parses metadata from the AI meeting note block (attendees, dates, action items), updates the properties on the meeting note page, and links it to the relevant records in our Contacts database. Others sync key data from apps like Xero (invoices) and Harvest (time tracking) into databases that we then built dashboards on top of. And another enables our Custom Agents to sign people up to our mailing list when they indicate interest in an email.
Notion’s legacy database automations feature hasn’t really evolved in years, and just like Zapier, making a big bet on AI agents and agentic coding as the future of automation for their platform.
🔥
Workers are free during beta until 11 August, after which they run on the same credits Custom Agents already use.
Once you sync external data into a Notion database with Workers, you can build pretty interactive dashboards
What this means for builders
No-code never really freed us from thinking like programmers. To build anything useful in a visual builder, we still had to understand conditional logic, branching paths, loops, how data passes from one step to the next, and what to do when a step failed. The drag-and-drop canvas only abstracted away the syntax; it gave us a way to translate that procedural know-how into working outcomes without writing the code ourselves.
AI coding agents let us do the same thing, just one step “closer to the metal”. The thinking is the same: we still reason about logic, sequence, and edge cases. What changes is that the agent writes and wires the actual code, so we get from idea to working automation faster and with a lot less manual assembly. Our attention moves further towards the business context, the ideas, and checking the result does what we intended, while the agent handles the busywork of stitching the logic together, testing, and troubleshooting.
Referral link; we earn a small commission if you sign up.
What if you don't build this stuff yourself?
None of this means you need to become a developer or a vibe coder. If you’re running a small team, it may still make more sense to find someone you can trust to set this up for you. The platforms are getting easier to use, but the architecture decisions (what to automate, what to keep deterministic, where to put the human checkpoints) still benefit from someone who has built this stack before.
On our side, the same shift means we can build and ship these systems for clients faster than ever. The time and effort to get from idea to a working, reviewed deployment has dropped sharply, and that benefit accrues to the teams we work with.
If that’s where you are, get in touch and we can map out what makes sense for your operation.
We built a contact enrichment pipeline for our Notion CRM using Apollo.io, HarvestAPI via Apify, and Zapier — for roughly $0–5/month. Here's the full architecture, the cost breakdown, and why CRM vendors charging $2–3K/year for the same capability is a broken model.
Lessons from building multi-turn AI agents across Zapier Agents, Notion Custom Agents, and Relevance AI — covering context window management, system prompt bloat, sub-agent trade-offs, and context rot.
Notion Custom Agents and Zapier Agents take fundamentally different approaches to AI automation. Here's where each one wins — and where it falls short.
