Agentic workflows invert the usual model: instead of you scripting every step, you give a model an objective and the tools it may use, and it plans and executes the steps itself, adapting as results arrive.
Scripted workflow vs agentic workflow
| Scripted workflow | Agentic workflow | |
|---|---|---|
| Who decides the steps | You, on the canvas | The model, at run time |
| Best for | Processes you already understand | Investigations and tasks you can only describe by goal |
| Predictability | High | Medium — depends on model quality and guardrails |
| Example | "Parse invoice → check total → file it" | "Investigate this failed device and tell me the likely cause" |
Both run on the same engine and appear in the same history.
Building an agent
- Open Workflow → Agentic.
- State the objective/instructions — what success looks like, in plain language.
- Grant tools — the MCP/external and local tools the agent may call. Agents are only as capable as the tools you give them; start minimal and expand as needed.
- Optionally attach a Deep Agent node inside a scripted workflow when only one step should be agentic.
Contain the blast radius:
Agents can act through their tools. Always include a User Approve node (or equivalent policy) before consequential actions — sending messages, changing configurations, deleting data. Start with read-only tools and widen access as trust builds.
Monitoring agents
While an agent runs, the run view shows its actions step by step — which tool it called, what came back, and its next decision. Use the telemetry/event-log panel to follow the reasoning trail, and review completed runs in History to audit every action taken. This trail is the basis for trusting (or correcting) an agent.
Where agentic fits in the platform
- Agentic admin capabilities — administrators control which agentic features are enabled for your workspace. See Agentic (Admin).
- Skills — package a tested procedure once, then let agents and workflows reuse it. See Skills.
- Tools — MCP servers and local tools agents call. See Tools (Admin).
Practical examples
| Goal | Objective sketch | Tools needed |
|---|---|---|
| Triage inbox incidents | "Read the incident description, gather related logs, summarise root cause, draft a reply. Do not send without approval." | Log search, knowledge lookup, email draft, User Approve |
| Vendor research | "Compare these five vendors on price, support, and certifications. Output a table with sources." | Web search, Output File |
| Config audit | "Check each device config against the policy checklist and list deviations." | Device read APIs (read-only), Format Check |
Best practices
- Describe success, not steps — let the model plan; over-specifying defeats the purpose.
- One objective per agent — chain multiple agents (or a workflow with a Deep Agent node) for multi-goal work.
- Budget and cap — long agentic runs consume tokens; check Usage regularly while tuning.
- Keep humans in the loop — approvals and Format Check gates before side effects.