AGENT WORKFLOW DESIGN
Agent Flow is a growing library of workflows built around one question: what happens when the agent hits something it can't resolve alone. Every flow has its handoff point defined before it ever runs.
Get early access →Early subscribers get the first flows before they're publicly available.
THE PROBLEM
Rolling out an agent for support tickets or lead qualification usually starts well — the routine cases go smoothly, the team is happy. Then an edge case shows up: a question outside the knowledge base, a mismatched invoice number, a reply nobody scripted for. The agent does something approximate, and doesn't say it's guessing.
The mistake surfaces two weeks later, in a complaint or a reconciliation error. The team loses trust and goes back to doing it by hand, now suspicious of automation as a category.
That's not a language model problem. It's the absence of a designed place for the agent to stop and ask.
DEFINITION
An agentic workflow is a business process mapped step by step, where an AI agent owns a defined share of the decisions and a human keeps the rest — by design, not by accident. It differs from classic automation in that the agent interprets content — an email, an invoice, a ticket — instead of replaying a fixed script. And it differs from a free-running agent in that its authority and its handoff point are settled before anything touches production.
THE METHOD
A script executes steps. A flow knows when to stop executing them.
Every flow in the library defines four things before the agent gets access to a single tool: what triggers it, which decisions it's allowed to make on its own, how to recognize the edge of that authority, and exactly who the case goes to when it's crossed — with full context, not just a "something went wrong" ping.
This isn't another comparison of agent frameworks — you'll pick the tooling based on whatever stack you already run. Agent Flow deals with the shape of the process, before anyone writes the first line of a prompt.
THREE FLOWS
The agent reads the inbound inquiry and enriches it with CRM data, sorting it into three buckets — ready for a call, needs clarification, outside the ideal customer profile. Sales only ever sees the first two. The third goes to the archive with a reasoning trail anyone can audit later.
Tickets reach the agent before a human sees them. It answers order-status questions, shipping timelines, and standard return policy on its own. The moment a customer asks for a refund above a set threshold, or the language signals escalation, the case moves to a person — with the full conversation history attached, not just the last message.
Invoices and delivery confirmations get read, matched against the purchase order, and posted automatically when the numbers match exactly. Any discrepancy past an agreed threshold — even a small one — routes to the accounting queue with the mismatch highlighted, instead of getting rounded away quietly.
WHO IT'S FOR
Agent Flow is useful before you pick a tool — as a process map any technical team or integrator can implement later. It's just as useful afterward, once the agent is live and nobody can quite explain what it does in the situations that don't fit the script.
QUESTIONS
No. We don't ship a runtime and we don't ask you to change your stack — LangGraph, CrewAI, n8n, or custom code all work. Agent Flow defines the shape of the process and the handoff point before anyone writes a line of code.
A flow is a decision map, not code — you can draft it yourself or with an integrator, then hand implementation to any dev shop or freelancer. That documentation is usually the missing piece in those projects.
RPA replays a recorded sequence of clicks and breaks whenever the interface or file format shifts. An AI agent reads the content and handles variation — but in return it needs explicit decision boundaries. In practice the two often run side by side: RPA executes, the agent decides. We cover the full comparison in Agentic workflow vs. RPA — what's the difference.
Mapping a single process usually takes days of workshops and writing, not months. The slow part isn't drawing the diagram — it's agreeing on the thresholds: above what amount, at what level of uncertainty, and to whom exactly the case escalates.
The first flows go out to subscribers over the coming months, in signup order. If you leave your email, we'll also suggest which process to document first for your industry.
EARLY ACCESS
Sign up to get the first flows before general release — along with a short conversation about the process you want to automate first.
No spam. We write when the flows are ready to download.