Agents & Workflows

"Agent" is the word of the year. Everyone is selling agents. But the difference between a workflow and a real agent is simple: who decides the next step — you or the model?

Workflow vs. Agent

• Workflow = predefined steps. "Read email → extract subject → summarize → send to Slack." The model does each step, but you wrote the order.
• Agent = goal + tools. "Answer this email", and the model itself decides whether to read the thread, search the web, check the CRM, or reply.

   WORKFLOW (deterministic):
   step 1 ──► step 2 ──► step 3 ──► done


   AGENT (autonomous):
                ┌──────────────┐
                ▼              │
   goal ─────► think ──────► tool?
                ▲              │
                └──────────────┘
            (loops until it says "done")

When to use which

• Workflow — when steps are clear, repetitive, auditable. E.g. processing 1000 invoices, generating weekly reports.
• Agent — when the task requires decisions along the way, exploration, or the path isn't known upfront. E.g. bug investigation, competitive research, code debugging.

Multi-agent: orchestrators and sub-agents

A "big" agent can delegate sub-tasks to specialized sub-agents. Like a manager splitting work across a team.

• Orchestrator = the main agent that receives the goal and splits tasks.
• Sub-agents = specialized agents (e.g. research, writing, validation). Each has its own context, its own tools.
• Benefit = separate context. Sub-agents don't pollute the orchestrator's context with their details.