Automation: Make unveils MakeOS, n8n keeps scheduled runs, UiPath tests live agents

This week's theme is control. Automation platforms are no longer asking only what an AI agent can do, but who watches it, who approves its steps and what happens when a server goes down. Make laid out its direction at Waves '26, n8n made scheduled runs more robust, UiPath now tests agents while they run, and Zapier made it simpler to connect AI tools.

Make: MakeOS and new rules for agents

On 20 October, at its Waves '26 conference, Make introduced MakeOS, an environment where agents, processes and actions are meant to be visible, permissioned and reversible. MakeOS is in closed beta; everyone else can join a waitlist. No pricing has been published. Existing scenarios keep working.

Part of the news is available now:

• agent orchestration: an agent can delegate tasks to subagents,
• Maia gets a Plan mode (you approve changes before they run) and memory across chats,
• Knowledge: shared company context for people and agents,
• Custom Roles and Private Spaces for finer permissions,
• Make Grid, a live map of objects and logs, plus audit logs for changes made by agents and MCP clients.

Announced but not yet available: pausing a scenario for human approval via Slack, Teams or the mobile app, tracking agent quality and AI spend, and organisation policies on which apps and models may be used. Process Intelligence (Task Mining and Process Mining, built on Celonis technology) is on the waitlist.

n8n: scheduled runs no longer get lost

n8n 2.45, released on 20 October, turns on the durable scheduler for all Schedule Trigger nodes with no extra setup. Each run is saved to the database before it is due, so a restart no longer drops pending jobs. A missed run still fires if it falls within its grace period. One behaviour changes: "every N minutes" schedules now count from when the workflow was published, not from the top of the hour. You can keep the old timing for now with N8N_SCHEDULER_TRIGGER_NODE_MODE=legacy.

The 2.46.0 beta from 27 October lets you use personal credentials in team projects without sharing them. Teammates without access to those credentials cannot run or publish the workflow with them, and cannot see the execution data.

UiPath: tests for agents in production

On 20 October UiPath released runtime evaluations in preview. A deployed agent is scored against datasets of inputs and expected outputs, on demand or on a recurring schedule. The aim is to catch regressions, accuracy problems and unexpected behaviour early. Real production runs can become new test cases.

Zapier: one MCP server for all your AI tools

In late October Zapier made it possible to connect Claude, ChatGPT, Cursor and other MCP clients to the same Zapier MCP server. They share apps, actions and skills, and every tool call lands in one history. In managed mode you choose which actions each client may run and can lock fields, such as a Slack channel. The feature is included in every Zapier plan.

What it means for your business

1. Add human approval to sensitive steps. Payments, quotes and data deletion should not be left to an agent without a check. The platforms are building this in, but you can set the rule today.
2. Check scheduled workflows after the n8n upgrade. If a report must start exactly on the hour and you use "every N minutes", switch to a cron or daily schedule.
3. Keep testing agents after go-live. A small set of examples with expected answers, run weekly, surfaces problems before your customers do.

If you want your workflows and agents set up so they stay under control, see our automation service or get in touch.