Workflow & no-code agents

Every hop in every workflow,decided.

Multi-agent pipelines fail one hop at a time. Zenstra Workflow Security checks each hop against policy and the task the workflow declared, before the next step runs.

The risk

One bad hop
and the whole flow follows.

Workflows chain agents, tools and MCP calls. An injected instruction at one step travels to the next.

  • Built outside ITNo-code flows wired into core systems without review.
  • Spreads between agentsInjected instructions passed along every hand-off.
  • Standing accessFlows running on keys with far more reach than any step needs.
A real pathOne email travels seven hops.
  1. 1intake-agentAn n8n intake agent receives a customer email carrying an instruction.
  2. 2hand-offThe instruction is passed along from agent to agent, hop after hop.
  3. 3hop 7The flow tries crm.export to an outside email address.
  4. BlockNot in the workflow’s declared task. The flow stops at hop 7.hop 7 · crm.export → outside email
Industry references
  • ASI07Insecure Inter-Agent CommunicationOWASP Top 10 for Agentic Apps
  • ASI08Cascading FailuresOWASP Top 10 for Agentic Apps
  • ASI01Agent Goal HijackOWASP Top 10 for Agentic Apps
Policy as code

How it is
stopped.

Each hop is checked against the task the workflow declared, so one bad step cannot carry the rest of the flow with it.

  • Decided inline
  • Fails closed
  • Versioned with every decision
denyhop.*when outside(workflow.task)denycrm.exportto outside emailholdcore.writefrom no-code flowsallowhandoffintake-agent → triage-agent
How Zenstra handles it

Map the flow.
Decide every hop.

Workflow Security sees the whole graph and decides each step before it executes.

Map

The workflow graph

Every agent, tool and MCP hop in n8n, Zapier, LangGraph and CrewAI, mapped and given an owner.

  • Every hop mapped
  • Owners for every flow
  • Unregistered flows flagged
Decide

Per-hop decisions

Each hop is allowed, held for a person or blocked before the next step executes.

  • Allow, hold or block per hop
  • Intent across agents
  • Next step waits for the decision
Record

The run, step by step

Every hop and decision is on the record, so a failed flow can be replayed and explained.

  • Replay any run
  • Policy version per decision
  • Tamper-evident record
Where it runs

Built for
the tools teams use.

  • n8n

    Self-hosted and cloud workflows with agent steps.

  • Zapier

    Automations that call AI and write to business systems.

  • Copilot Studio

    Agents built by business users inside Microsoft 365.

  • LangGraph

    Graph-based agents your engineers build and run.

  • CrewAI

    Multi-agent crews with hand-offs between roles.

  • MCP

    Tool calls made from any step in the flow.

FAQ

Common questions.

Do we need to change our workflows?

No. Workflow Security works with the flows you have; there is no agent rewrite.

What does a held hop look like?

The workflow pauses at that step until a person approves or rejects it, and the decision is recorded.

Does it see agent-to-agent hand-offs?

Yes. Each hand-off is checked against the original task, so an injected instruction cannot ride along.