AI Agent Approvals, Safety & Monitoring
AI Agent Approvals, Safety & Monitoring
Agents run without anyone watching, so DataMagik keeps a person in charge of every change they make and records everything they do. This page covers how to control what an agent may change, how changes are approved, who gets told, and where to see what happened.
What an agent may change
Two settings in the agent's About this agent panel decide this, and both apply:
- What it can use — for each group of tools, Read lets the agent look things up and Can change lets it make changes. With every Can change off, the agent only reads.
- Approval & safety — how much it may do without a person:
| Autonomy | What happens when the agent wants to change something |
|---|---|
| Only reads | It looks things up and reports. It never changes anything; any write is refused. |
| Asks before changing anything (default) | It files a proposal for a person to approve. The run waits as Waiting for approval and continues from that step once approved. Nothing is changed if it is rejected. |
| Acts on its own | It changes things without asking. Everything it does is still recorded. This needs a plan add-on; without it the option is greyed out with "Acting on its own is not part of this plan, so this agent will ask first." |
Choose Who approves its changes in the same section. This is a role: anyone who holds it can decide the agent's proposals. If it is left as Nobody yet, a change the agent wants to make is refused rather than queued, and the Workbench shows a banner asking you to choose an approver.
Approving an agent's proposals
Proposals arrive in the approvals inbox of everyone holding the approver role (the envelope icon in the top bar — see Approvals). Each shows the agent's name, what it wants to do and a line of evidence. Press Review to open it; an email the agent wants to send shows Review & send and can only be approved from its review page, never in bulk. You can also reach proposals from a run's Review this run's proposals, or from the link in a Teams, Slack or email notification — the link takes you to DataMagik to sign in; nothing is ever approved from the message itself.
The review page says "An AI agent proposed a change. Nothing happens until you approve it." and shows:
- The change — for a script fix, a before/after comparison (with Open in script editor); for an email, the recipients, subject and body. You can edit the recipients under Send to, and must fill one in if the agent did not know who to send to.
- What the agent saw — press Show the evidence to see what led to the proposal.
Press Approve & apply to make the change, or Reject (optionally with a reason). If what the proposal changes has moved on since it was made, approving marks it Out of date and changes nothing. Agents can never approve their own proposals.
Telling people (notifications)
Under Tell people in the agent's settings, choose Who to tell — one or more of the company's notification channels — and When to tell them:
- It proposes a change that needs approval — one message per run, with links to review.
- A run finishes with something to report.
- A run fails — for unattended runs, with the reason.
Press Save who to tell. Channels — Email, Slack and Microsoft Teams — are set up by a security administrator under Admin → Notification channels, where Send test checks a channel works.
Agent runs
The Runs tab (also Admin → AI Assistant → Agent runs) lists every agent run, what it did and what it cost, with totals for the last 7 days. Filter by All / Failed / Waiting / Succeeded or search by agent name; History on the Agents list opens it filtered to one agent.
Open a run to see what started it (Run now, Scheduled, Condition, Automation, Event, Escalation, Inbound message, Chat, Workbench or Step test), the model it used, its input and output, each step, the tools it called, its findings, the proposals it filed and their decisions, anything that Needs a person, and its tokens, cost and duration. A run's status is one of Queued, Running, Waiting for approval, Succeeded, Failed, Timed out, Cancelled or Skipped. A failed run says why, and a failed step stops the rest of the run.
Save as fixture keeps a run as a recorded check (see Fixtures below). Runs are kept for 90 days. Model usage also appears in AI Token Usage.
Stopping every agent (kill switch)
At the top of the Runs page, Agent execution shows whether agents may run — normally "Running. Agents may call their allowlisted tools." A security administrator (DataMagik - Security) can press Stop agents, giving a reason, to stop every agent in the company: it then reads STOPPED for this company and no agent tool call runs from the next call on. Resume agents undoes it.
Memory
Agents can remember what they learn. The Memory tab shows every note, which agent learned it, the evidence, and how established it is — Seen once, Seen by separate runs, or Checked by a person. Confirm a note so future runs trust it, or Withdraw a wrong one with a reason; every run that used a withdrawn note is flagged Needs review. Notes that are not seen again fade and expire after 90 days.
Fixtures
A fixture is a recorded check: it replays an agent's saved version with its tools stubbed, so it calls no tool and no model and costs nothing. Fixtures are captured automatically when you Save as agent in the Workbench, or from a run with Save as fixture. On the Fixtures tab press Run or Run all to replay them; a failed fixture shows each step's expected and actual result. Catalog agents replay their fixtures before installing or updating.