cutover-community
Blog
November 17, 2026

No Dark Agents: AI That Shows Its Work, on the Model You Choose

An AI agent that fixes a problem and leaves no evidence has not made you more resilient. It has given you one more thing to explain to an auditor.

That is the idea behind Cutover’s 2026.21 release. The AI agent audit trail is now part of how agentic tasks run: the agent reads what was already recorded, does the work, and logs its own evidence, without touching the instructions it was given. Cutover AI can also now fill in the Runbook Homepage, the page everyone outside the working team watches for status, so nobody has to stop mid-incident to type an update. And connecting your own models over MCP is simpler. In major incident management and IT disaster recovery, the record of what happened matters as much as the fix.

Cutover is the orchestrator of record for humans and agents, at enterprise scale. This release is about the “of record” part.

In the open, on the record

The agent to worry about is the dark one. It wants all of your data, tells you little about what it did, and may be training on what it sees. Regulated operations need the opposite: every agent action visible, attributed, and approved, on a model you chose.

A runbook already works that way for people. The task description is the instruction. What happens during the run is evidence. Agents should add to that evidence, and when the plan itself needs to change, that goes through the same approvals and audit trail as it would for a human.

That agent does not have to be ours. Some banks are ruling out SaaS AI altogether and want to run their own model. Agentic tasks support custom AI agents as well as the built-in Cutover AI agent, so your approved model can work inside the same governed runbook.

In 2026.21, the built-in Cutover AI agent in an agentic task can read what was already recorded on a task and add comments of its own. You instruct it in plain English:

“Check the latest comments for the deployment URL and post the test results as a comment.”

The agent then:

  1. Inspects prior task notes for context.
  2. Completes the actions.
  3. Logs its findings and evidence as comments, for your review and approval.

The description stays as it was approved. The AI agent audit trail builds up beneath it, run after run.

Why the AI agent audit trail matters most in tech ops

Code is getting easier to prove. Tech ops is not.

A recovery or a major incident spans applications, infrastructure, and people. You cannot prove it end to end. You can only run it, then show what happened. That makes the run record the closest thing to proof that operations has.

That record is not a system log. Logs show what the systems did. The runbook shows how the work was done: the tasks, the instructions, who or what acted on them, and where execution deviated from the plan. None of that appears in a log, and the how is what people and agents need to learn from.

Foundation models know what is written. Cutover captures what is done. If agents are going to do real work inside a run, they have to write to that record the way a human resolver does: attributed, in sequence, open to review. An agent that works off the record weakens the one artifact your regulator will ask for.

Less copy and paste, more deciding

Every incident and every failover ends with the same chore. Someone copies what happened into fields so the dashboards and reviews make sense. The Runbook Homepage is where everyone outside the working team looks for status, so when its fields are stale, more people interrupt the ones doing the work.

Cutover AI now does that directly:

  • The Cutover AI Assistant and the Cutover AI agent can write into custom fields on the Runbook Homepage.
  • The Cutover AI Assistant can populate PIR custom fields once your runbook is completed.

The people in the room stay on the information that matters in the moment. The review fields get filled while the detail is fresh, not from what someone remembers on Friday afternoon.

Bring your own model, with less friction

Many of our customers do not want another AI model to approve. They want the models their security teams have already cleared to become useful in operations. The Cutover MCP server exists for exactly that.

Two developer updates make it easier:

  • Simplified MCP setup. Clearer URL guidance and a default API base URL reduce connection errors when you connect a compatible AI client to Cutover. Setup instructions, examples, and full documentation are in the public repo and on the Cutover MCP Developer Portal page.
  • User context for external systems. Shipped one release earlier, in 2026.20: you can use the {{User.unique_id}} dynamic variable in integration and data source configurations. When Cutover sends a request to an external system, it swaps in the current user’s unique ID, so the other side knows who is acting. For security reasons it works only in integrations and data sources, not in runbook or task fields such as titles or descriptions.

The principle is the same throughout. When work crosses a system boundary, identity and evidence go with it.

The fair objection: is this just more noise?

A skeptic would say comments are not control. An agent that narrates its work can still do the wrong thing. A long thread written by a model is one more thing for a tired Major Incident Manager to read at 3 a.m. And a tidy record can breed false confidence: evidence of activity is not evidence of correctness.

All fair. Three things answer it.

  1. The record is there to be reviewed. Agentic tasks keep humans in the loop. The agent logs its findings and evidence for your review and approval.
  2. Changes stay governed. Progress goes into comments, so the approved description is not overwritten as a side effect. If the plan should change, that is a decision, made with approval and on the record.
  3. The alternative is worse. With no record you have neither control nor evidence. With one, you can check, challenge, and improve.

A record does not make an agent right. It makes an agent checkable. In regulated operations, that is the bar.

Also in this release

  • Bulk return to draft. Admins can return up to 100 templates or snippets to draft status in one go from the workspace.
  • Context for approvers. Requesters can add comments in the Request Approval modal. Reviewers see them in the approval flow and in notification emails, and they are logged in the audit trail and CSV reports.
  • A more careful Assistant. The Cutover AI Assistant now checks custom field values before sending them. If a value is not valid, it corrects itself or tells you what is wrong.

The full list is in the release notes.

Frequently asked questions

Does the Cutover AI agent change my task descriptions when it logs progress?

No. In an agentic task it can read existing comments and add its own to record progress and evidence, without overwriting task descriptions.

Which fields can Cutover AI fill in?

The Cutover AI Assistant and the Cutover AI agent can write into custom fields on the Runbook Homepage. The AI Assistant can also populate PIR custom fields when a runbook is completed.

How do I connect my own AI client to Cutover?

Through the Cutover MCP server. Setup now has clearer URL guidance and a default API base URL. Start with the Developer Portal page.

Where can I use the user dynamic variable?

In integration and data source configurations only. It cannot be used in runbook or task fields such as titles or descriptions.

Make your agents show their work

Agents are coming to your operations either way. The question is whether they leave a record you can stand behind.

Cutover AI needs to be enabled on your instance, so existing customers should speak to their Customer Success Manager. If you are new to Cutover, see it for yourself: cutover.com/book-a-demo.

Ky Nichol
CEO
AI
Product
Major incident management
IT disaster recovery
Latest blog posts