Make AI Agents: 6 Workflows for DevOps & Security

See the workflow in action, tap through the tabs below:

You triage inbound checkout-service alerts.
Classify severity using the runbook tool.
Only page on-call for High or Critical, after
checking the last 24h for duplicate alerts.
Never modify infrastructure directly.

New alert: {{payload}}
Check if this matches a known false positive
pattern. If it's real and Sev1/Sev2, page
on-call with a one-line summary and the
relevant runbook link. Otherwise log it and
stop.

A dependency bump PR just opened for
{{package}} {{old_version}} to {{new_version}}.
Summarize what changed, flag anything security
relevant, and note if we've had issues with
this package before.

Make has been the "connect anything to anything" tool for years, 3,000+ app integrations, a visual canvas, routers, filters, the works. In February 2026 they shipped something different: Make AI Agent (New), an app that doesn't just move data between systems, it reasons about what to do with it.

What Make AI Agents Actually Are

Here's the distinction that matters. A normal Make scenario is deterministic. Trigger fires, steps run in order, same input always gives the same output. An AI agent is goal driven. You give it a job description, a pile of tools (which are just other Make scenarios), and it decides which tool to call and in what order based on whatever request just landed in its lap. You can watch it think, too. Make's Reasoning panel shows the agent's step by step decisions right on the canvas, which is the difference between "trust me" and "here's exactly why I paged you at 2am."

Don't confuse this with Maia, Make's other AI feature. Maia is the co-worker that helps you build scenarios inside the editor, you chat with her, she writes the automation for you. The AI Agent app is what runs after you've built something, the thing that actually makes decisions in production. Use Maia to build faster. Use Agents to handle judgment calls your existing scenarios can't.

Quick Setup

You don't need to touch code. Everything happens in the same builder you already use for scenarios.

1. In your Make navigation panel, go to AI Agents and click Create agent.
2. Name it and write a description of its job and its limits. Something like: "You triage inbound security alerts. Classify severity, check for known false positive patterns, and only page on-call for High or Critical. Never modify infrastructure directly."
3. Connect a language model. Free and Core plans can use Make's own AI provider with no separate account. Paid plans can also bring a custom connection (OpenAI, Anthropic Claude, Gemini) with your own API key, which is usually the cheaper option once you're running agents regularly.
4. Click Create agent. It's live but useless until you give it tools.

The Mindset: Treat It Like a Fast Intern, Not an Oracle

Make's own docs put it well: pick tasks for your agent that you'd trust an intern to handle on their first week. Categorizing tickets, drafting a summary, researching a CVE, sure. Approving a production deploy or touching billing, no.

The reason this matters more for DevOps and security than for, say, marketing copy, is that your tools are the blast radius. An agent with a "restart service" tool and bad judgment will restart the wrong service with complete confidence and a very reasonable sounding explanation. Scope the tool list tight. Give it read access and drafting power generously. Give it write access and destructive power sparingly, and usually behind a human approval step.

Also worth knowing going in: agents don't remember anything between runs. Every trigger is a fresh brain with no memory of the last ticket it triaged. That's a feature for consistency, not a bug, but it means your instructions and knowledge files have to carry all the context every single time.

Six Workflows Worth Stealing

1. Alert triage and on-call paging. Point a webhook or Datadog/PagerDuty trigger at an agent whose only tools are "look up runbook," "check if this alert fired in the last 24 hours," and "page on-call via PagerDuty." See the ALERT TRIAGE tab above for the exact prompt. You get triage that doesn't wake anyone up for noise, with full reasoning logged for the postmortem.

2. Dependency and CVE research. Trigger off a Dependabot or Renovate pull request. Tools: "fetch package changelog," "search internal wiki for prior incidents involving this package," "post PR comment." See the CVE RESEARCH tab above for the prompt. The agent does the fifteen minutes of changelog archaeology nobody wants to do before every merge.

3. On-call handoff notes. Run this on a schedule, end of shift, every day. Tools: "pull last 24h of resolved incidents from your tracker," "pull relevant Slack thread summaries." Prompt: "Summarize the last 24 hours for the incoming on-call engineer. Group by severity, note anything still being watched, and flag open questions." Nobody starts a shift blind again.

4. Access request triage. Feed it Slack or Jira access requests. Tools: "check current role in identity provider," "check if requested access matches a pre-approved policy," "post to approvals channel." Prompt: "A user requested {{access_level}} to {{system}}. If it matches a standard, pre-approved role for their team, approve and log it. Otherwise route to the security channel with your reasoning." Cuts your approval queue down to the requests that actually need a human.

5. Log anomaly research. When your anomaly detector fires, hand the agent the raw signal plus a "query internal docs" knowledge base and a "run diagnostic query" tool. Prompt: "This metric spiked: {{anomaly_details}}. Check our internal runbooks and past incident notes for similar patterns, then summarize likely causes ranked by probability." This is the workflow where the agent's reasoning panel earns its keep, you can see exactly which prior incident it pattern matched against.

6. Security ticket bulk triage. Point it at your ticket queue. Tools: "classify ticket type," "assign to team," "tag priority." Prompt: "Classify this ticket as vulnerability report, access request, or policy question. Assign priority based on our SLA rules and route to the right team channel." Great for the Monday morning backlog nobody wants to open first.

Gotchas and Safety

Make AI Agent (New) is still in open beta, which means pricing and behavior can shift under you. Don't build anything mission critical on it without a fallback path.

Every tool scenario needs a Return output module at the end, or the agent gets nothing back and can't finish its job. This trips up almost everyone on their first build.

Keep the tool list short. Five tools an agent understands well beats fifteen it gets confused about. If the agent keeps picking the wrong tool, that's usually a tool description problem, not a prompt problem, rewrite the description, not just the instructions.

And treat any agent touching production systems the way you'd treat a new hire's access, least privilege by default, read-heavy, write-light, and a human in the loop for anything destructive. The reasoning panel is great for debugging after the fact. It is not a substitute for guardrails before the fact.

Credits and Cost Tips

Running an agent through Make's built in AI provider costs one credit per operation plus token based credits for whatever the model burns. Bring your own OpenAI or Anthropic API key on a paid plan and it drops to a flat one credit per operation, you pay the model provider separately. If you're running agents constantly, that split usually wins on cost.

Uploading knowledge files (PDFs, docs, spreadsheets) for an agent to reference costs embedding tokens once at upload time, then query time costs after that. Don't re-upload the same runbook every time it changes, update it in place.

Plans start at the Free tier (1,000 credits a month, no agent access) and move up through Core, Pro, and Teams, each priced around 10,000 credits a month with the per-credit rate dropping as you go up. Check the official Make pricing page for current numbers since this is a beta feature and pricing is explicitly subject to change.

FAQ

Is a Make AI Agent the same thing as a regular scenario?
No. A scenario runs the same steps every time. An agent gets a goal and a toolbox of scenarios, then decides which ones to call and in what order based on the specific request it receives.

How much does it cost to run one agent workflow?
It varies with model size and token usage, but expect a handful of credits per run using Make's built in provider, less if you connect your own API key on a paid plan. Test with "Run once" before you turn anything loose in production to get a real number.

Do Make agents remember past conversations?
No, by design. Every run starts fresh with no memory of prior runs. Whatever context matters has to live in your instructions, your knowledge files, or the input itself.

Closing

Make AI Agents won't replace your on-call rotation and they shouldn't run your deploys. But the boring reasoning work, is this a false positive, has this package burned us before, what happened on last night's shift, that's exactly the kind of judgment call a well scoped agent handles cleanly. Start with one narrow workflow, watch the reasoning panel for a week, then decide what earns a second one.

Want the systematic version of this instead of another tool tutorial? Our automation and AI courses walk through the whole workflow-design process, and if you're comparing low-code agent platforms, see how it stacks up against n8n's approach to DevOps and security workflows.