Your engineering team spends too much time on work that doesn’t require human judgment. Bug triage, PR review checklists, deploy verification, incident categorization — these are patterns, not problems. AI can handle them.
What Are Agentic Workflows?
An agentic workflow is a defined sequence of AI-powered steps that executes autonomously — with human checkpoints where they matter. Unlike a simple “ask the AI and get an answer” pattern, agentic workflows:
- Execute multi-step processes — research, analyze, decide, act
- Use tools — read files, query APIs, create tickets, post messages
- Handle branching logic — different paths based on AI analysis
- Report results — structured output for human review
A Real Example: Automated Triage
Here’s a workflow I built with SWEny for automated issue triage:
- Ingest — new issue appears in GitHub/Linear/Sentry
- Analyze — AI reads the issue, related code, recent changes, and similar past issues
- Categorize — severity, component, likely root cause
- Enrich — add relevant context, link related issues, suggest assignee
- Route — assign to the right team with a summary and recommended priority
This runs on every new issue. The team reviews the AI’s categorization instead of doing it themselves. Time saved: ~15 minutes per issue, across dozens of issues per week.
Design Principles
DAGs, not chains. Model your workflow as a directed acyclic graph, not a linear chain. Steps that don’t depend on each other can run in parallel. Steps that fail can retry independently.
Checkpoints for trust. Start with human approval at every decision point. As you build confidence, remove checkpoints one at a time. Never go from “human reviews everything” to “AI does everything” in one step.
Observability first. Log every step, every AI decision, every tool call. When (not if) something goes wrong, you need a complete trace.
Idempotent steps. Every step should be safe to retry. If the workflow fails halfway through, you should be able to restart from the failed step without side effects.
Getting Started
You don’t need to build an orchestration engine from scratch. Start with one workflow that automates your team’s most repetitive task. Define the steps, build the integrations, add checkpoints, and iterate.
If you want a head start, SWEny is open-source and purpose-built for this. Or reach out and I’ll help you design workflows for your specific team.
For more on AI workflow patterns and the Claude Agent SDK, visit nateross.dev.