Docs / Running it
Running Mindset without spending hours on it
One named person, a backup, and about ten minutes a day. Most of the watching is done for you.
What runs without anybody asking
The first of the four is the one people do not expect.
- Every Mindset agent carries behaviour criteria, written when it was built, describing what it must do and must not do.
- A new version goes live only if every single criterion passes. Not most of them.
- The set that passed is recorded against that version, so months later you can see what a given version was held to.
The criteria are written by whoever built the agent, so the review that matters is of the criteria themselves. That goes to the system owner, once, before the first version goes live.
What none of it does is notice on your behalf that an agent which ran every morning for six weeks stopped on Tuesday. That is what the two habits above are for.
Asking without signing in
Neither habit means learning new software. The guide to the Mindset record is itself an agent, published to your team over MCP, so the question is asked wherever they already work and the answer comes back in the same window.
The same route produces the evidence.
- Ask for a month of approvals as a table, and paste it into a board pack.
- Ask which agents hold write operations and on what, and hand the answer to an auditor.
- Or send the full record over OpenTelemetry into whatever your team already retains, and never ask at all.
What intelligence Mindset offers for different security, ops & leadership teams
Security, operations and leadership look at the same record and arrive with different questions. Every one is answered from what was recorded, and every one can be asked from Teams, Copilot, Claude or ChatGPT.
If you are responsible for security
Your question is what an agent could have done, and what it did.
| Question | How often |
|---|---|
| Which agents hold write operations, and on what? | Monthly, and after any change |
| What was each agent granted, and what did it actually call? | Weekly |
| Is anything running under a person's credentials? | Monthly |
| What was approved, by whom, and how quickly? | Weekly |
| Show me every action in this run, in order | During an investigation |
| Which runs did this run hand off to other agents? | During an investigation |
| Get the raw traces into our own tooling | Set up once |
The one people do not think to ask is the second: what each agent was granted, against what it actually called.
- Granted and never called means a stage is being skipped in practice.
- Called and never granted is worth looking at the same day. It usually means a run delegated to another agent that came with resources of its own.
If you run the operation
Your question is whether the agent is working.
| Question | How often |
|---|---|
| Did the agent run at all? | Daily |
| Did the agent run as often as it should have? | Daily |
| How far did the agent get? | Daily |
| What is the agent waiting for? | Daily |
| Calls, failures, successes, average time, last used | Weekly |
| What is working, failing, idle or never used? | Weekly |
| Which version ran? | When something changed |
Idle and never used are separate states on purpose. Something that has never run has a trigger problem. Something that used to run and stopped has a different problem.
If you are paying for Mindset
Your question is whether the spend was worth it.
| Question | How often |
|---|---|
| Total spend, and how it moved | Monthly |
| Average cost per call | Monthly |
| What does each agent cost? | Monthly |
| What are we paying for capability we may not need? | Quarterly |
| How is spend split across suppliers? | Quarterly, before a negotiation |
| Who is using it? | Monthly |
| How much of the spend is cached, and what would the same work cost without caching? | Monthly |
| How many agents are live, and how many people use each one? | Quarterly |
Two numbers Mindset cannot give you. Time returned against the manual baseline, and how many people can now do something only one person could do before. Both come from your own systems, and the baseline has to be captured before the agent goes live because nobody can reconstruct it afterwards.
On caps and exports. A cost cap is set on the model connection, alongside the model, a token ceiling, PII handling and data residency, rather than on each agent. Set a cap on anything scheduled before leaving that agent alone for a month. There is no CSV download of runs. If you want the data in your own tooling, that is the OpenTelemetry export.
The four jobs, and how often each comes up
Permissions attach to agents rather than people, so this is not an access matrix.
Build it. Anyone. They know the process being automated, which is the part that cannot be delegated.
Decide which operations exist. Whoever is accountable for the system being touched. A Mindset operation is one specific named thing an agent may do on one system: get the purchase order matching this invoice number, rather than access to the finance system. The system owner is asked once which of those should exist, and that list is the boundary for every agent built against that system afterwards.
Approve writes. A named person in the function that owns the outcome. Finance approves finance writes, not the AI team. A Mindset owner approving writes across every system becomes a bottleneck and a rubber stamp at the same time.
Watch it. One named Mindset owner and a backup. The only standing commitment.
The cadence
| When | Who | What |
|---|---|---|
| Daily, ten minutes | Mindset owner | What failed, what is waiting on an approver, what ran and changed nothing. |
| Weekly, thirty minutes | Mindset owner and backup | New agents and new operations from the week. What each agent was granted against what it actually called. |
| Monthly, an hour | Add security and a system owner | Cost by agent and model. Agents that have not run in thirty days. Anything your discovery tooling found with no sanctioned route. |
| Quarterly | Add leadership | Time returned against baselines. Owners and backups reconfirmed. Agents archived. |
Put the monthly hour in the diary before the first agent goes live. The failure mode is not that the review is hard. It is that nobody schedules the review, and the first one happens after an incident.