Agentic workflows for the product leader
Choosing a workflow pattern, knowing when more than one agent earns its cost, and what it takes to make a workflow durable enough to be worth owning end to end.
A workflow adds two problems a single agent run does not have: choosing a shape for many steps, and surviving the real time a long process takes.
A single agent's loop only gets you so far. Some jobs are too big for one context window, too parallel for one worker, or too long-running to fit inside one continuous run — they need to pause for days waiting on a human, survive a restart, and pick up exactly where they left off. "Agentic workflow" is the name for what happens once a task needs more than one loop, or needs the loop to keep its place across real time. Getting this right is mostly a small number of orchestration and durability decisions, made deliberately instead of by whichever framework happened to be closest to hand.
A note on scope. Several neighbouring lessons already cover parts of this topic. Multi-agent systems & protocols develops the orchestration topologies and the MCP and A2A landscape. AI agents covers the workflow-versus-agent line, and the run-level controls: budgets, approval gates and the kill switch. Flowable builds a real process engine from scratch. Agentic AI as a product covers workflow capture as a business strategy. This module does not repeat them. It adds the decisions that sit between them: which workflow pattern to use, when more than one agent earns its cost, and what makes a workflow survive real time.
Where the depth lives
Each lesson summarises an idea and points to the lesson that covers it in full.
| If you want the full depth on | Read |
|---|---|
| Orchestration topologies and the MCP and A2A landscape | Multi-agent systems & protocols |
| The autonomy dial, and workflow versus agent | What an agent is |
| Run limits, approval gates and the kill switch | Running an agent in production |
| Writing the prompts inside a chain | Prompt chaining and multi-step workflows |
| Wait states, retries, compensation and process engines | Flowable |
| Workflow capture and agent economics | Agentic AI as a product |
The knowledge graph
Three questions asked in sequence
Which shape, how many agents, and will it survive real time?
Lesson 1Choosing a workflow pattern
Chain · route · parallelize · orchestrate · evaluate
Five patterns. Pick the first one that does the job.
Lesson 2Orchestrating more than one agent
Orchestrator and subagents
When a subagent earns its cost, and how to brief one.
Lesson 3Making it last
Surviving pauses & restarts
Saved state · waits · retries · undo
Owning the whole workflow
The moat — one step vs. end-to-end
Read it as three questions in order. Which shape? Pick the simplest of five patterns that does the job. How many agents? Add one only when it clears a named bottleneck. Will it last? Once the shape exists, it must survive a three-day wait and a restart. Only then is owning the whole workflow worth asking.
The lessons
- Choosing a workflow pattern — the five patterns, what each costs, and how to choose the simplest that works.
- Orchestrating more than one agent — when subagents are worth their cost, what the evidence says, and how to write a brief.
- Making a workflow durable, and worth owning — saved state, waits, retries, safe repeats and undo, then the case for owning the whole workflow.
Each lesson has a 🎯 For the product leader briefing, a labelled worked example, an "Under the hood" section for engineers, and a Sources list.
📌 Close out the module: Recap & real-world examples, which ends with a self-test.