Running a playbook and an agent side by side: ownership per alert category

You will run playbooks and agents together for a while, so assign each alert category an owner. Vendors document different autonomy models: per alert routing, four modes, three modes. This guide shows how to read those claims and set ownership by category with a rollback.

The question is who owns the alert, not which technology wins

The "after SOAR" argument often skips the middle. Most SOCs will run deterministic playbooks and agents in the same queue for a year or more. The practical question is ownership: for each alert category, who acts first, who can act without a person, and who is accountable when it goes wrong.

Vendors answer this differently, and the differences matter more than the AI label. Here is what they document, with the claim status marked.

Four ways vendors split the work

Per-alert routing. Swimlane says Turbine routes each alert to a deterministic playbook, AI-assisted investigation or fully agentic investigation, chosen per alert by intelligent routing (Swimlane home page) (read 2026-11-03, vendor claim). Coverage of its January 2026 launch says agents can replace several deterministic steps inside a playbook (SiliconANGLE, 2026-01-27) (read 2026-11-03). Ownership sits with the router, so the router's rules need an owner too.

Four autonomy modes. D3 says Morpheus offers Deterministic, AI-Assisted, AI-Led and Autonomous modes, and that alerts the AI cannot resolve pass to a human. It also says migrated playbooks arrive in Deterministic mode (D3 Security) (read 2026-11-03, vendor claim). That gives a clean ladder: move a category up one mode at a time.

Three autonomy modes with escalation written in. Mate documents Recommend only, Supervised and Auto-run under guardrails and audit. Its Gamebooks define when an agent must escalate, stop or ask for approval (Mate platform page) (read 2026-11-03, vendor claim). Mate says autonomy is raised one action class at a time and that a category drops back to supervised if accuracy drifts. Mate also says it tests a Gamebook against past cases and runs it in shadow mode before sign-off (Mate Gamebooks) (read 2026-11-03, vendor claim). Gamebooks reached general availability on 2026-08-26, per trade press (SiliconANGLE, 2026-08-26) (read 2026-11-03). Mate announced a 35M USD Series A on 2026-07-28 (Mate newsroom) (read 2026-11-03). Mate is not a SOAR, and we score it on what it documents, so it has no playbook editor to compare here.

Autonomous or human on the loop. Torq says response can run fully autonomously or with human-on-the-loop oversight, with a manual override and audit logs. Its agents run on a context model that records verdicts, exceptions and overrides (Torq home page) (read 2026-11-03, vendor claim). Torq's figure of over 90 percent of cases closed autonomously is reported by third parties and was not on the page we read, so we do not use it.

One more pattern is worth knowing. Splunk's agents sit at build time. The Automation Builder agent in Cloud release 8.6.0 builds playbooks, and the Connector Builder Agent in 8.7.0 builds connectors. They do not decide cases (Splunk SOAR Cloud release notes) (read 2026-11-03). If you want a run-time agent, that is a different product decision.

Treat accuracy numbers as claims until the method is public

Several vendors publish figures. D3 says up to 95 percent of alerts reach a graded verdict or an L2-depth investigation in under two minutes. Mate says 98.7 percent accuracy on 7,200 alerts and 97 percent false-positive auto-closure. Swimlane says Hero AI cuts MTTR by 75 percent. None of the pages we read gives the sample source, labelling method, period and who measured. For Mate, the sample size is stated in a chart caption and the measurer is the vendor. Our rubric does not score these figures above 8.0 on the AI criterion without that method. Run your own numbers on your own alerts, and score them with your own labels.

A method for assigning ownership by category

Do this with a spreadsheet and two weeks of history.

List categories by volume and by blast radius. Phishing triage, EDR malware detections, identity anomalies, cloud misconfigurations, vulnerability notices. Rank each by alerts a week and by what a wrong action costs.
Mark the safe actions. Enrichment, tagging and ticket creation are reversible. Disabling an account or isolating a host is not trivial to undo. Give agents the first group early and the second group late.
Pick the owner for each cell. For every category and action class, write one of: playbook, agent with approval, agent alone, human. One owner per cell.
Set the fallback. If the agent fails, times out or declines, what runs? The default should be the existing playbook or the human queue, never silence.
Log both. Capture playbook outcome and agent outcome on the same alert for the same period. That comparison is your evidence.

Start with the category that has high volume and low blast radius. Phishing triage is the usual first choice, and it is a choice, not a rule.

Rules that keep the two from fighting

One writer per field. If the playbook sets severity, the agent proposes and does not overwrite.
One closer per alert. Two systems closing the same case is how audit trails break.
A shared kill switch. You need one place to stop agent actions that does not also stop the playbooks.
Names in the log. Every action should record the actor, playbook or agent, and the version.
A review cadence. Weekly at first: sample closed alerts from each owner and re-label them by hand.

What to ask a vendor

Ask five questions. At what level is autonomy set: per action, per category or per product? What happens to an action the agent proposes and the guardrail blocks? Can the vendor show the log of one decision with its evidence? What is the documented limit of the agent, in the vendor's own words? How do you turn it off for one category without turning it off for all? Google's documentation of Gemini playbook creation is an example of stated limits: no playbook blocks, no parallel actions, only installed integrations (Google SecOps, create and edit a playbook with Gemini) (read 2026-11-03). Most agent pages do not state limits like that. Ask for them.

Checklist for a side by side run

List alert categories with weekly volume and the cost of a wrong action.
Classify each action as reversible or not.
Assign one owner per category and action class.
Define the fallback when the agent fails or declines.
Set one writer per field and one closer per alert.
Build a kill switch for agent actions only.
Log actor, version and evidence on every action.
Run in shadow or recommend mode first, with analysts labelling outcomes.
Review a weekly sample by hand.
Move a category up one autonomy step only after a written review.