What is Suggest / Recommend / Act?
Suggest, Recommend, Act is the three-step policy spectrum that governs what any consumer of an Agentic Operations Center is allowed to do. At Suggest, the consumer may investigate and describe a fix. At Recommend, it may stage the exact change for a human to approve. At Act, it may execute within a bounded scope, with the action recorded. The same spectrum applies to NeuBird, to an engineer at the console, and to an agent a platform team wrote last week. Movement along it is earned by track record, which is why NeuBird calls the model earned write access.
The three levels
Suggest: the consumer queries telemetry in place, reasons over engineered context, and returns a conclusion with the evidence cited and a proposed fix described in words. Nothing changes. Recommend: the consumer stages the exact change (the kubectl command, the Terraform diff, the runbook step) and a named human approves or declines it; the approval is recorded. Act: the consumer executes changes of a specific, pre-agreed kind within a bounded blast radius, and every action lands in the audit trail with the evidence that justified it. Nothing executes at any level without an approval the audit trail can show, whether that approval was given per action or per class of action in advance.
| Level | What the consumer may do | Human role | Recorded |
|---|---|---|---|
| Suggest | Investigate, conclude, describe a fix | Reads and decides | Conclusion and evidence |
| Recommend | Stage the exact change | Approves or declines | Change, approval, outcome |
| Act | Execute within a bounded scope | Sets the scope, reviews the trail | Every action with evidence |
Why the same spectrum for every consumer
Most organizations grant permissions per tool and per identity, so a vendor agent, an engineer, and a homegrown agent end up with three unrelated permission models and no single place to answer the question a CIO asks after an outage: who was allowed to do that? Putting every consumer on one spectrum, governed by the center rather than by each agent, produces one policy, one write boundary, and one audit trail. It also lets trust be earned: an agent that has recommended the right fix for a class of incident a hundred times can be promoted to Act for that class, and demoted if it is wrong once.
What to remember
- 1Suggest, Recommend, Act is the policy spectrum for every consumer of the center, human or agent
- 2Nothing executes without an approval the audit trail can show
- 3Movement along the spectrum is earned by track record, per class of action
- 4One spectrum for every consumer means one policy, one write boundary, one audit trail
Frequently asked questions
Is Suggest the same as read-only?
NeuBird avoids the label. Suggest is observe-only in effect, but it is the first step on a spectrum the consumer can move along, not a fixed permission grant.
Who decides when an agent moves from Recommend to Act?
The organization, per class of action, on the basis of the agent’s recorded track record for that class. The center enforces the decision and records every action taken under it.
Does Act mean fully autonomous?
It means autonomous within a bounded, pre-approved scope. The approval was given in advance for that class of action, and every execution is recorded with its evidence.
See it in action. No slides.
NeuBird AI compresses incident investigation from hours to minutes: autonomous root cause analysis, with zero manual triage.