# Agentic Workflows: Design One on a Page and Run It by Hand > Agentic workflows without new software: map the job in a six-column sheet, run it as a prompt chain in ChatGPT or Claude, and know when to move on. _Source: https://professionalstoolkit.com/articles/agentic-workflows — The Professional's Toolkit · updated November 2026_ _Part of: [Run a multi-step job with AI agents that hand work to each other](https://professionalstoolkit.com/solutions/run-agentic-workflows) — the guide, the do-it-yourself version and the tools that do it_ --- **Agentic workflows** are multi-step jobs that an AI carries from start to finish: it works through the steps, uses tools, checks its own output and stops at set points for a person to approve. You don't need new software to run your first one. Write the job down as steps on a single page, then run those steps by hand as a chain of prompts in the assistant you already have. That manual version suits a weekly job with four to six steps. It stops working when steps must hand off on their own, run side by side, retry or leave a log. **6 columns** — Step, input, tool, output, who checks, on failure: the whole design of a workflow fits in one sheet - **4-6 rows** — the size of a workflow you can run by hand - **3 runs** — clean manual runs before you automate anything - **5** — active scheduled tasks on ChatGPT Plus, per OpenAI's help page ## What agentic workflows are, and what they are not An ordinary prompt asks for one answer. An agentic AI workflow breaks the job into steps, and each step takes the previous step's output as its input. Anthropic, the maker of Claude, draws a useful line here: workflows are systems where models and tools follow predefined paths, while agents direct their own process and tool use.[1] For a first project, the predefined path is what you want. It is predictable. You can read it. That is the agentic workflow meaning worth keeping: a written procedure in which a model does the work inside each step and a person owns the checkpoints. An [AI agent](/articles/what-is-an-ai-agent) is the single worker; the workflow is the line of work around it. The wider idea, including where autonomy fits, is covered in our guide to [what agentic AI is](/articles/what-is-agentic-ai). The same Anthropic guidance recommends the simplest solution that works, with complexity added only when needed. A sheet and a chat window are the simplest solution. ## Design the whole job on one page Open a new Google Sheet and give it six columns: Step, Input, Tool, Output, Who checks, On failure. Then write one row per step. Most jobs fit in four to six rows. If you need twelve, you are describing two workflows. Fill the columns in this order: 1. **Step.** A verb and an object: *collect competitor updates*, *draft the brief*. One job per row. 2. **Input.** Exactly what the step receives: a list of URLs, last week's brief, the previous row's output. 3. **Tool.** What the step needs beyond writing: web search, a file, a spreadsheet, or nothing. 4. **Output.** The format you expect back, down to headings or columns, so the next step can use it without cleanup. 5. **Who checks.** A name, or *nobody*. Any step that sends, books, publishes or writes into a system of record gets a name. 6. **On failure.** What happens when the output is wrong: rerun with a note, fall back to last week's version, or stop and tell someone. The last two columns are the ones people skip, and they are what separates a workflow from a long prompt. Anthropic calls the check between steps a gate: a test on an intermediate result that confirms the process is still on track before the next step builds on it. In a manual run, you are the gate. Writing down what you look for now is what lets software, or a colleague, do the looking later. > 💡 **Write the output format first:** Fill the Output column before you write any prompt. A step whose result has fixed headings or columns can feed the next step without cleanup. ## Run it by hand as a prompt chain Prompt chaining means each prompt processes the output of the one before it. The trade is speed for accuracy, because every call gets an easier task. A project workspace in your assistant keeps the chain's context in one place. In ChatGPT, the steps are these:[2] 1. Select **New project** in the sidebar and name it after the job. 2. Open the more options menu, choose **Project settings**, and add project instructions: the audience, the house format, the rules that hold for every step. 3. Add the reference files the steps need: PDFs, spreadsheets, docs or pasted text. 4. Start a chat for row one. Paste the row's input and ask for the row's output format. 5. Check the result against your *Who checks* column. If it passes, paste it into the next prompt as the input for row two. 6. When a result is worth reusing, open the message menu on it and select **Add to project sources**. Claude works the same way. A project there is a self-contained workspace with its own knowledge base and project instructions, and free accounts can create projects too.[3] Whichever assistant you use, run the checking step in a fresh chat instead of the one that wrote the draft, so the reviewer judges the text and not its own earlier reasoning. Then add a second tab to your sheet and log each run: the date, which step failed, what you changed. That tab is your run history. > 💡 **Keep one prompt per row:** Save each row's prompt in your sheet next to the row. When a step fails you edit one prompt, not the whole chain, and the next run starts from the corrected version. ## Agentic workflow examples: a weekly competitor brief Take a job many small teams do late or not at all: a Monday brief on what three competitors changed last week. As a workflow it is four rows. | Step | Input | Tool | Output | Who checks | On failure | |---|---|---|---|---|---| | Collect updates | three competitor sites | web search | table: change, date, link | you open each link | rerun for the missing company | | Draft the brief | the checked table | none | one page, three headings | nobody | rerun with a note | | Check the draft | draft plus table | none | claims that have no link | you | back to the draft step | | Send | the approved brief | email | sent message | you, every time | hold until tomorrow | The value sits in the third row. A model asked only to *find claims with no source* has a narrow job and does it carefully, which is the whole logic of chaining. The same four-row shape covers other recurring work. An inbound lead: research the company, draft a reply, approve, send. A supplier invoice: pull the fields, compare them with the order, flag the differences, approve. A customer review: classify it, draft a response, approve, post. In each one the last row carries a person's name. > ⚠ **The send step always has a name:** Agents make mistakes. Anything that is sent, booked, published or written into a system of record needs a person's approval first, in a manual chain and in any tool you move to later. ## Scheduled tasks stretch the manual version one step further One built-in feature is worth trying before you add any tool. ChatGPT can run one-time or recurring tasks on a schedule, and eligible paid plans can start a task from Gmail, Slack or GitHub activity.[4] So the first row of the brief can run itself every Monday. Read the limits first. OpenAI's help page says a task created in a project cannot access files uploaded to or stored in that project, the number of active tasks depends on the plan (5 on Plus), and Free accounts get flexible windows such as morning instead of exact times. Notice what a scheduled task is: one prompt on a timer. It does not pass its result to a second step, wait for your approval and then start a third. You still carry the output forward yourself. For a weekly brief, that is fine. For work that arrives all day, it isn't. ## What breaks first when you run the chain by hand The manual chain breaks at specific points, and your run log will name them. - **Hand-offs.** Each step waits for you to copy and paste. If the job should start when an email arrives, or it runs more than a few times a week, you are the bottleneck. - **Parallel steps.** Twenty companies to research means twenty chats, one after another. - **Retries.** When the third step fails overnight, nothing reruns it or falls back. - **Logs.** Your second tab records what you remembered to type, not what happened. One sign alone is rarely a reason to switch. Two usually are. The tools built for this differ mainly in who builds the workflow and how. The three below cover the common situations, and each still expects a person at the risky step. If your sheet has several rows with a name under *Who checks*, and nobody on the team writes code, a visual canvas is the closest match to what you already drew. Each row becomes a node, and each gate becomes a setting on the line between two nodes. Relevance AI does this with [a multi-agent canvas called Workforce](/articles/relevance-ai-workforce). ### Relevance AI — Your sheet redrawn as a canvas Workforce is a canvas of trigger, agent, tool and condition nodes. A hand-off either always passes the work on or lets the agent decide, and each connection can require approval or cap automatic runs before it asks a person. The Tasks page collects approvals, escalations and errors. - Where it stops: Connections between agents are one-way. Standard tools and agents stop after 15 minutes. The builder runs only in a desktop browser. - Plan: Workforce is on every plan: Pro, Team and Enterprise. There is no free plan. - [Build a multi-agent canvas](https://professionalstoolkit.com/articles/relevance-ai-workforce) · [Relevance AI](https://professionalstoolkit.com/tools/relevance-ai) When the pain is volume instead of approvals, the twenty-company problem, look for a tool that splits one task among helpers working at the same time. That is where [Gumloop's agents, connectors and triggers](/articles/gumloop-ai-playbook) fit: operations and marketing people whose bottleneck is a long list of similar items, not the approvals between steps. ### Gumloop — Parallel subagents for long lists An agent can start up to 10 subagents at once for one task, each with its own tools and context. Scheduled, app and webhook triggers start the work, approvals arrive in chat or Slack, and Evaluations grade finished tasks. - Where it stops: A trigger turns off after 3 failed runs in a row. When credits run out, agents stop and trigger runs fail. - Plan: Pro. There is no free plan, only a one-time trial of Pro that needs a card. - [Connectors, triggers and credits explained](https://professionalstoolkit.com/articles/gumloop-ai-playbook) · [Gumloop](https://professionalstoolkit.com/tools/gumloop) Developers have a different lane. If someone on the team is comfortable in Python and wants the workflow kept in version control, a code framework keeps every step, condition and checkpoint in files the team can review. CrewAI's own guidance is to use a Flow when you control the steps, their order and the conditions, which is exactly what your sheet describes. Our guide to [CrewAI Flows](/articles/crewai-flows) shows when to use one and how to build it. ### CrewAI — The workflow as Python code Flows are event-driven steps with shared state, conditions and branching, and a human-feedback step pauses a run for approve, reject or revise. Checkpointing saves state so a run can resume after a failure. The hosted CrewAI AMP platform adds tracing. - Where it stops: The framework needs Python skills and your own LLM provider key. On AMP, Basic allows 2 automations and 50 workflow executions a month. - Plan: The open-source framework has no plan. CrewAI AMP: Basic (free) or Enterprise. - [When and how to use Flows](https://professionalstoolkit.com/articles/crewai-flows) · [CrewAI](https://professionalstoolkit.com/tools/crewai) These are three of several routes. Genspark, for instance, lets you describe a workflow in plain language and run it on a schedule or when an email arrives. The parts a full setup needs (roles, orchestration, memory, tools, approval and monitoring) and which product covers which part are laid out in our guide to [multi-agent systems](/articles/multi-agent-systems). ## FAQ **How is an agentic AI workflow different from ordinary automation?** Ordinary automation follows fixed rules: when this happens, do that. In an agentic AI workflow a model does judgment work inside the steps, such as researching, drafting or classifying, so the results vary and need checkpoints that a rule-based automation does not. **How do you build an agentic workflow without code?** Write the steps in a sheet, run them as a prompt chain in a ChatGPT or Claude project, and log each run. When you need automatic hand-offs, move the same design to a no-code tool with a visual canvas or plain-language setup. **Do agentic workflows need more than one agent?** No. One model working through well-defined steps is enough for most first workflows. Separate agents start to pay off when steps need different tools or instructions, or when many items must be processed at the same time. **What is the difference between an agentic workflow and an AI agent?** In a workflow the model and its tools follow a path you wrote down in advance. An agent decides its own steps and tool use as it goes. A first project is easier to check as a workflow, because every step and every checkpoint is on the page. ## Start with one sheet and three clean manual runs Pick one recurring job this week and write its rows before you open any tool. Run it by hand three times and log every failure. If the log stays short, keep the manual chain and put a scheduled task on the first row. If it shows the same waiting, the same copy and paste or the same missed rerun each time, you now hold a tested design to carry into a workflow tool. You will also know, early in any trial, whether that tool can run your sheet. ## Sources [1] Anthropic — https://www.anthropic.com/engineering/building-effective-agents (Nov 2026) [2] OpenAI Help Center — https://help.openai.com/en/articles/10169521-projects-in-chatgpt (Nov 2026) [3] Claude Help Center — https://support.claude.com/en/articles/9517075-what-are-projects (Nov 2026) [4] OpenAI Help Center — https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt (Nov 2026)