Security teams can’t keep up with AI-driven risks by manually working violations one at a time. Coding agents ship code at machine speed. Frontier AI models are shrinking the time between discovery and exploitation. The tech debt that was tolerable at human speed is now exposure on a clock. Keeping pace takes agent-driven action: analysis and remediation that start the moment risk appears.
Cycode Maestro and Agentic Workflows already combine to analyze exposure, conduct complex security investigations, and orchestrate multi-agent workflows that confirm exploitability and drive remediation. Cycode has extended this to facilitate organization-specific analysis and automation with two new powerful capabilities:
- Custom Maestro agents. Build your own Maestro agent to perform the analysis your team does manually today, from upgrade impact analysis to credential validity to code ownership mapping. Define what the agent should analyze, the evidence it should weigh, and the outcomes that decide what happens next.
- Custom Maestro agents in Agentic Workflows. Add any custom agent as a step in an Agentic Workflow. It runs automatically on every matching violation, and its outcome determines the next action: open a fix, route to the owner, escalate, or pause for human approval.
Together, they turn your team’s security expertise into human-controlled and agent-driven action that runs at scale.
Create Custom Maestro Agents to Automate Security Tasks
Security teams frequently repeat the same analysis on many violations: determining whether a package upgrade introduces breaking changes, tracing who owns a repository, weighing whether a finding touches production. Custom Maestro agents take that work on. On every run, the agent receives the violation that triggered it, investigates the code and context around it, and returns a verdict with a justification and a confidence score.
Building an agent takes four steps in the AI agents library:
- Describe it. Name the agent and give it a short description so anyone in your organization can find and reuse it.
- Define the analysis. In plain language, tell the agent what to evaluate and the evidence to weigh, such as upgrade impact, file paths, and repository names.
- Set the outcomes. Define up to three outcomes, each with a description the agent uses to decide when it applies. Maestro adds a reserved Failure outcome automatically, so a case the agent can’t resolve always has a path.
- Test it on a real violation. Run the agent against one of your own violations and see the verdict, its reasoning, and the tokens the run consumed. You know how the agent behaves, and what it consumes, before you publish or add it to a workflow (more on that below).
Custom agents are shared across your organization. Each one appears in the library alongside the built-in agents, with a count of the workflows that use it.
Add Custom Maestro Agents to Agentic Workflows
Analysis pays off when it drives action. In addition to creating custom Maestro agents, users can also add them to Agentic Workflows that start from a violation trigger.
When the agent runs, the workflow follows each outcome defined when the custom agent was created. Under each branch, you choose what happens: open a Jira ticket, message a Slack channel, set severity or status, add a label, open a fix pull request, or hand off to another agent for a second stage of analysis. Custom agents work across every violation category and act on new violations the moment they’re detected.
Control is built into every step:
- Confidence thresholds. An outcome branch runs only when the agent’s confidence meets the threshold you set, 70% by default. Raise it to act only on the clearest calls.
- Human approval where required. Turn on “Require human approval” for the first action in any outcome branch. The agent decides, a reviewer sees its recommendation in Required Approvals, and nothing happens until they approve.
- Stable logic. While a workflow uses an agent, its outcomes are locked, so the branches you built stay valid as the agent’s instructions evolve.
- Predictable consumption. Token consumption for every run is attributed to its workflow on the AI usage dashboard, and a ready-made template alerts you when usage reaches 80% of your quota.
Put Agentic Security to Work Across Your Program
Because custom Maestro agents follow your instructions and your outcomes, they adapt to your organization’s thresholds, teams, and environments. Here are a few of the ways teams can put them to work.
Burn Down Tech Debt Without Breaking Changes
Risk tells you what’s dangerous. It doesn’t tell you how much effort it requires to fix or whether upgrading a vulnerable dependency will introduce breaking changes. An Upgrade Impact agent analyzes target violations and classifies them by whether they require engineering effort to fix, and if so, how much effort.
Run the workflow retroactively across your backlog, then keep it on for every new violation. No-effort fixes open a fix pull request. Engineering changes can be passed to downstream agents to generate non-breaking fixes. Inconclusive verdicts get a label and a notification. Quick wins clear themselves, and engineering capacity goes to the fixes that need engineers. That is frontier AI readiness in practice: less exposure waiting in the queue.
Coordinate Fixes for Widespread CVEs
When a vulnerability lands in a widely used package, dozens of teams can end up with near-identical tickets for the same problem. A widespread or isolated CVE agent counts open violations with the same CVE across your organization, against a threshold you set. Widespread CVEs get one coordinated ticket and a Slack alert. Isolated ones go to the owning team. When AI-driven discovery produces a surge of new disclosures, that is the difference between one coordinated response and a flood.
Separate Live Secrets from Test Credentials
A leaked credential and a placeholder in a test file can look identical. An agent that analyzes the file path, repository name, and surrounding code tells them apart. Real credentials go straight to your incident channel and the repository owner. Test or sample values get a label and an ignored status, with configurable human approval on the status change.
Prioritize and Route by Business Context
Some of the most valuable agents bring business context to every decision, so each violation reaches the right team with the right urgency:
- Owning team: Infer ownership from code owners, the repository’s purpose, and the affected code, then route to that team’s board and channel.
- Production IaC: Distinguish production deployments from dev, staging, and sample configuration, and raise severity where it matters.
- Externally exposed code: Escalate violations reachable from public APIs or user-facing input.
- Sensitive data handling: Escalate violations in code that processes PII, payment data, or credentials.
Security as Agentic as Development
Developers are using coding agents to ship faster. Attackers are using frontier AI to find weaknesses faster. Cycode equips security with AI advantages to combat AI risk.
Automate the analysis only your organization can define, test it before it goes live, and let workflows act on every outcome with confidence thresholds and human approval wherever you want it.
Custom Maestro agents and their use in Agentic Workflows are available now for Cycode customers. Request a demo to get started building your first custom agent and running it in your environment.