On 30 November 2026 OpenAI switches off Agent Builder, the drag-and-drop workflow canvas it launched only a year earlier. At the same time n8n's valuation has reportedly doubled to $5.2 billion and Zapier, Make and n8n are all rebuilding themselves as AI agent orchestrators. This guide explains what is happening in AI workflow automation, how the platforms differ, and how to choose between a no-code tool and code-based agents built on Claude, OpenAI or Mistral.
Why AI workflow automation is being reshuffled now
For a decade, workflow automation meant "if this, then that": a new form entry creates a CRM contact, an invoice email lands in a folder. Large language models changed what a single step can do. A step can now read a contract, classify a support ticket, draft a reply or decide which branch a workflow should take. That turns automation platforms into the place where companies actually run AI in production, and it explains why both the automation vendors and the AI labs are fighting for the same budget.
OpenAI retreats from no-code: what the Agent Builder shutdown means
OpenAI introduced AgentKit in October 2025, with Agent Builder as a visual canvas for multi-agent workflows. Many commentators read it as OpenAI's entry into the market of Zapier, n8n and Make. On 3 June 2026 OpenAI announced that it is winding down Agent Builder and its Evals platform. According to OpenAI's deprecations page, existing evals became read-only on 31 October and both products disappear on 30 November 2026. ChatKit stays. OpenAI's recommended paths are the code-based Agents SDK or ChatGPT Workspace Agents, and for evaluations the open-source tool Promptfoo.
| Date | OpenAI platform change | What to do |
|---|---|---|
| 26 Aug 2026 | Assistants API shut down | Responses API and Conversations API |
| 31 Oct 2026 | Evals become read-only | Export results, move test sets to Promptfoo or your own harness |
| 30 Nov 2026 | Agent Builder and Evals dashboard/API shut down | Agents SDK (code) or ChatGPT Workspace Agents |
The automation platforms become agent orchestrators
While OpenAI steps back, the established platforms are moving forward. All three have added AI agent steps, LLM nodes and support for the Model Context Protocol (MCP), so an agent in Claude or ChatGPT can trigger their workflows as tools.
- Zapier is the largest by reach, with connectors to more than 8,000 apps, and is famously capital-efficient: it has raised very little outside funding. AI steps and agents bill as tasks on the same plan as other Zaps. It is the easiest start for business teams, but every step counts toward the bill.
- Make (part of Celonis) offers a visual scenario builder with more control over branching and data mapping than Zapier, and it bills per operation. It is popular with agencies and operations teams that build complex, multi-branch scenarios.
- n8n is the fastest-growing of the three. It raised a $180 million Series C at a $2.5 billion valuation in October 2025 (total funding $240 million), and in May 2026 its valuation was reported at $5.2 billion after a strategic investment from SAP. Its key difference: it can be self-hosted, it bills per workflow execution instead of per step, and it lets developers drop into JavaScript or Python inside a node.
Sources: n8n Series C announcement (Oct 2025); May 2026 figure as reported after SAP's strategic investment.
Investors are betting that the "orchestration layer" between AI models and business systems will be worth a lot, whichever model wins. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% in 2025. The same firm also warns that over 40% of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear value or weak risk controls. Both numbers point to the same conclusion: the tooling is ready, but most failures come from picking the wrong process or the wrong architecture.
Source: Gartner press releases (June and August 2025). 2026 and 2027 values are forecasts.
Where Anthropic, Mistral and Manus fit
The AI labs approach automation from the other side: instead of a canvas with boxes, the model itself plans the steps and calls tools.
- Anthropic (Claude) never shipped a visual workflow builder. Its bet is code and protocols: the Claude Agent SDK for building agents in Python or TypeScript, Claude Managed Agents for hosted, long-running agents, routines in Claude Code for scheduled jobs, and MCP, which it created and donated to the Linux Foundation, so the same tools work across Zapier, n8n, Make and custom code.
- OpenAI now points developers to the Agents SDK and business users to Workspace Agents inside ChatGPT: code for engineering teams, chat-native agents for everyone else.
- Mistral offers an Agents API with built-in connectors and open-weight models that can run on your own servers. For European companies that need data to stay on-premises, a self-hosted n8n plus a self-hosted Mistral model is a fully sovereign stack (see our sovereign AI guide).
- Manus is a general-purpose autonomous agent: you describe a goal and it plans and executes the work. It suits ad-hoc research and one-off tasks better than repeatable, audited business processes (more in our Manus guide).
No-code platform or code-based agents? A comparison
There is no single winner. The right choice depends on who builds and maintains the workflow, where the data may live and how many times a month it runs. The radar below is our assessment from client projects, on a 1 to 5 scale.
Assessment by AI Workshop / Blocksize based on project experience, November 2026. "Code SDK" means agents built on the Claude Agent SDK, OpenAI Agents SDK or Mistral Agents API.
| Option | Best for | Trade-off |
|---|---|---|
| Zapier | Business teams, fast start, widest app catalogue | Per-step billing gets expensive at volume; cloud only |
| Make | Complex visual scenarios, operations teams, agencies | Per-operation billing; cloud only |
| n8n | Technical teams, GDPR-sensitive data, high volumes | Self-hosting means you run updates, backups and security |
| Code SDK (Claude, OpenAI, Mistral) | Core processes, deep agent logic, full testing and version control | Needs developers; you build monitoring and retries yourself |
| Vendor builder (e.g. Agent Builder) | Prototypes only | Lock-in and deprecation risk, as 30 November shows |
The hidden cost driver: how each platform counts
Pricing models differ more than list prices. Zapier counts a task for every action step, Make counts every module run as an operation, and n8n counts one execution per workflow run, however many steps it has. For a 10-step AI workflow, that is a tenfold difference in billable units.
Illustration of billing models, not a price comparison: per-step = runs × 10 steps; per-run = one unit per run. Model API costs for AI steps come on top in every option.
Model tokens are a separate line. An agent step that sends a long document to Claude or GPT costs the same whether Zapier or n8n triggers it, so the main levers there are model choice, prompt caching and keeping prompts short.
A practical roadmap for 2026–2027
- Audit Agent Builder and Assistants usage now. Anything still built on them must move to the Agents SDK, Workspace Agents or another platform before 30 November.
- Start with one boring, high-volume process. Invoice intake, ticket triage and lead enrichment pay back fastest. Measure time per case before and after.
- Let business teams prototype in no-code, then harden. When a workflow becomes business-critical or high-volume, move it to n8n or code with tests, logging and version control.
- Use MCP as the contract between layers. Expose your systems as MCP tools once; Claude, ChatGPT and your automation platform can then all use them.
- Keep humans in the loop for actions with consequences. Payments, contract changes and customer emails need an approval step and an audit log, and agents need least-privilege credentials (see AI agent security).
The bottom line
AI workflow automation is consolidating around two layers: orchestration platforms like Zapier, Make and n8n, and code-based agent SDKs from Anthropic, OpenAI and Mistral, joined by MCP. OpenAI's exit from visual building is a reminder to keep workflows portable. For most companies in Croatia and the DACH region, the pragmatic stack in 2026 is a no-code tool for quick wins, n8n or code for core and data-sensitive processes, and a model choice that can change without rebuilding everything.
Which of your processes should an AI agent run first?
AI Workshop maps your workflows, picks the right platform for each one, from Zapier and n8n to Claude-based agents, and helps you move anything that depends on OpenAI Agent Builder before the shutdown. Our Claude Certified Architects build automations that are tested, monitored and portable.
Talk to an AI consultant