What did Make add to AI Agents on October 9, 2026?
Make added two features to the Make AI Agent app on October 9, 2026: a fallback AI provider connection that takes over when the primary provider or model is unavailable, and prompt caching that lets the agent reuse its instructions and tool definitions. Both are available on all plans. The same release added Claude Sonnet 5.5 and GPT-6.1 Sol and a way to improve agent answers with Maia.
The source is Make's release note Make AI Agent improvements, new AI models, SSO login routing, and MCP tools for audit logs, updated October 9, 2026. A day earlier, on October 8, Make also launched event subscriptions, which can alert you when a scenario run fails.
| Feature | Date | Plans |
|---|---|---|
| Fallback AI provider connection for Make AI Agent | October 9, 2026 | All plans |
| Prompt caching for Make AI Agent | October 9, 2026 | All plans |
| Improve agent responses with Maia | October 9, 2026 | All plans |
| Claude Sonnet 5.5 and GPT-6.1 Sol models | October 9, 2026 | In the Make AI Toolkit, provider and Make AI Agent apps |
| MCP tools for audit logs (read-only) | October 9, 2026 | All plans |
| Event subscriptions (signed alerts for events such as a failed run) | October 8, 2026 | Teams (up to 3 per organization) and Enterprise (unlimited) |
How does the Make AI Agent fallback connection work?
You set a second AI provider connection on the agent. If the primary provider or model fails or is unavailable, Make retries the run once using the fallback connection. If you do not choose one, Make's own AI Provider is the default fallback.
What happens when the AI provider fails
That single retry is the important limit. A fallback covers a provider outage or a model that is briefly unavailable. It does not cover a bad prompt, a tool that returns an error, or a rate limit that hits both providers. Pick a fallback model that can follow the same instructions and call the same tools, and test the agent on it before you rely on it, because two models can give different answers to the same prompt.
For a lead-handling agent, the practical setup looks like this: primary model from one provider, fallback from a different provider, and an alert if both fail so a person can follow up by hand. The alert part is where event subscriptions come in.
What does prompt caching do in Make AI Agents?
Prompt caching lets the agent reuse its instructions and tool definitions instead of sending them in full on every call. Make says this reduces latency and credit usage. For supported Anthropic Claude models you choose a cache duration; supported OpenAI models reuse prompt content automatically.
One detail from the release note matters for chat-style agents: with Claude models, the cached context applies across repeated calls in the same conversation when you provide a conversation ID. If each run starts a new conversation, you will get less benefit. Agents with long instructions and many tools, called often, gain the most. A short prompt called a few times a day gains little.
Make changed how it prices AI tokens on August 25, 2026, separating input and output token rates. Caching cuts the repeated input part of each call, which is usually most of an agent's prompt.
How do Make event subscriptions alert you when a scenario fails?
An event subscription sends a real-time notification to a URL you choose when a subscribed event happens, such as a scenario being created or a scenario run failing. Each delivery is signed with a secret so your system can check it came from Make, and each subscription keeps a log of every delivery attempt.
Event subscriptions are on the Teams plan, with up to three per organization, and on Enterprise, with no limit. For a business that runs leads or orders through Make, this replaces checking the scenario history by hand. A failed run can post to Slack, open a task in your CRM or text the person on call. If you use GoHighLevel, the alert can go to an inbound webhook that creates a task in the right sub-account; our guide to GoHighLevel webhooks and integrations shows how that connection works.
What does this change mean for businesses running AI agents in Make?
It removes two common reasons an AI agent quietly stops working: a provider outage and nobody noticing the failed run. With a fallback connection and a failure alert, an agent that answers leads or sorts requests keeps running through most short outages, and a person hears about the rest.
What we would set up on a production agent now (our Make.com automation agency page covers credits, error handling and builds):
- Fallback from a different provider than the primary, tested with the agent's real tools.
- Prompt caching on agents with long instructions, with a conversation ID where the agent holds a conversation.
- An event subscription for failed runs (on Teams or Enterprise), sent to the place your team already watches.
- A human fallback in the workflow itself: if the agent cannot answer, the lead goes to a person with the context attached.
Zapier made a similar move this month when it moved Agents into the Zap editor, and n8n is adding its own controls in version 2.43. For an AI agent that answers leads end to end, see the dental case study, or our AI agents overview.
These are the same rules we use for AI agents in GoHighLevel and n8n. See our AI workflow automation service, or compare platforms in our n8n and Zapier pricing research.
Questions about the Make AI Agent update
Is the Make AI Agent fallback connection free?
The feature is available on all Make plans. The fallback run still uses credits or provider tokens like any other run. Check your provider's pricing for the fallback model, since it may differ from your primary.
Which AI models work with Make prompt caching?
Make's note says supported Anthropic Claude models let you choose a caching duration, and supported OpenAI models reuse prompt content automatically. It does not list every model, so check the Make AI Agent app settings for your model before you count on savings.
Can I get event subscriptions on the Make Core or Pro plan?
No. As of October 8, 2026, event subscriptions are available on the Teams plan, with up to three subscriptions per organization, and on the Enterprise plan, with unlimited subscriptions. On other plans, use error handlers inside each scenario to send alerts.
Sources
- Make Help Center, Make AI Agent improvements, new AI models, SSO login routing, and MCP tools for audit logs, updated October 9, 2026, read October 12, 2026.
- Make Help Center, Introducing event subscriptions, updated October 8, 2026, read October 12, 2026.
- Make Help Center, Release notes 2026 (AI Provider token pricing change, August 24, 2026), read October 12, 2026.
