# An agent can fix a narrow set of findings without a person, and the set depends on the artifact it produces

Guide, 26 Oct 2026, read 10 Nov 2026

**The short answer.** Vendors describe four outputs an agent can produce on its own: a pull request, a patch job, a network or web application firewall rule, and a ticket with the work scoped. Each fits a different finding type. Vendor pages read on 10 Nov 2026 show which outputs each platform states.

## Four outputs, four kinds of finding

Automated vulnerability remediation is a loose phrase. The useful question is what the agent hands over. The vendor pages read for the register describe four outputs, and each applies to a different kind of finding.

A dependency, container image or infrastructure-as-code finding in a repository can be fixed by a pull request. An operating system or third-party application finding on a managed host can be fixed by a patch job. A finding with no patch, or in a system nobody will touch this quarter, can be reduced by a firewall or web application firewall rule. Everything else arrives as a ticket, ideally with the plan attached.

| Output | Finding type it fits | Platforms stating it (vendor's claim, read 10 Nov 2026) |
| --- | --- | --- |
| Pull request | Dependencies, container images, infrastructure as code | Averlon, Backline, Rival |
| Patch job through a tool | Operating system and third-party application patches | Astelia, Qualys, Tenable, Vicarius |
| Network or web application firewall rule | Findings with no patch, or no change window | Astelia, Rival, Seemplicity, Zafran |
| Ticket with plan | Anything that needs a person to schedule it | Onit, Rapid7, Seemplicity, Zafran |

## What the vendor pages say about each output

Averlon says its GitHub Actions generate ready-to-merge pull requests for container images, Terraform and Pulumi, and Helm charts, with a ticket per pull request assigned to the code owner. Backline says each remediation includes a plan and a ready-to-review pull request for dependency, container image and infrastructure findings, and that existing unit tests run before delivery. Rival describes a flow of plan, test, and open a pull request for human review.

On patch jobs, Qualys says its TruRisk Eliminate 2.0 chooses between zero-touch, ring-staged and manual change board deployment by patch reliability, and rolls back automatically if a deployment causes issues. Tenable says Patch Management applies policy-based patching that can be scheduled, paused or rolled back. Vicarius lists patching on demand, scheduled or by rule across Windows, macOS, Linux and over 20,000 third-party applications.

On rules, Astelia describes a drafted firewall rule waiting in an approval queue. Zafran describes a compensating control, such as a web application firewall rule, applied in a few clicks at the choke point of an attack chain. Seemplicity added Response Options on 2 Sep 2026, giving compensating controls and configuration changes alongside the full patch.

## Where a person is still in the loop

Every vendor above keeps a person somewhere in the path. Averlon's own example expects a developer to review each generated fix, at about 15 minutes per fix in its example. Backline says fixes ship as standard pull requests, so rollback uses existing Git workflows. Onit states that nothing executes without approval. Cogent states that high-impact changes require human approval and an audit trail.

So the working question is not whether a human is involved, but at which point and for how many items. A reviewer who approves 260 grouped pull requests is doing a different job from one who approves 110,000 findings (the Averlon example, vendor's claim: 3,400 tickets became 260 pull requests).

## Findings an agent should not touch alone

Platforms differ in how they handle the last of these. Onit states that when evidence is missing an agent says unknown with the reason instead of guessing. Check that your platform of choice does the same, and that unknown items land in a queue a person reads.

- A patch to a system with no tested rollback.
- A change to an authentication or network path that many services share.
- A fix that requires an upgrade across a major version of a library.
- Any finding where the platform reports unknown owner or unknown exposure.

## Sort your own backlog by output first

Before reading a profile, take 100 open findings as a sample and sort each by the output that would close it: pull request, patch job, rule, or ticket. The mix tells you which platform output matters to you. A backlog that is mostly container image and dependency findings points to pull requests. One that is mostly operating system patches on managed hosts points to patch jobs.

Most backlogs hold a mix, and the mix changes the question. If 60 of the 100 need a pull request, a platform with strong patch tooling and no repository access leaves most of your queue untouched. The profiles state the outputs each vendor describes, and the none found labels show where a page was read and the capability was not on it.

## Related pages

- [Guides](/guides/)
- [How approval should work when an agent changes production](/guides/how-approval-should-work/)
- [How to measure a vulnerability backlog in hours, not findings](/guides/how-to-measure-a-backlog/)
- [Ownership mapping: finding who can fix a finding](/guides/ownership-mapping/)

## Sources

- Averlon blog on automating remediation: https://www.averlon.ai/blog/automating-remediation-application-infrastructure-security-issues (read 10 Nov 2026)
- Backline, product page: https://backline.ai/product/ (read 10 Nov 2026)
- Qualys, agentic AI page: https://www.qualys.com/apps/agentic-ai (read 10 Nov 2026)
- Tenable Patch Management: https://www.tenable.com/products/patch-management (read 10 Nov 2026)
- Seemplicity, Response Options, 2 Sep 2026: https://seemplicity.ai/news/response-options/ (read 10 Nov 2026)
- Onit Security, platform page: https://onit.security/platform/ (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://automatedvulnerabilityremediation.com/guides/what-an-agent-can-fix-without-a-human/
