Workflows

A single agent handles a single job. Most real work is several jobs in sequence, where one result feeds the next: research something, draft from the research, review the draft, then send it.

Workflows let you wire agents together into one unit that runs on demand or on a schedule.

Find them under Workflows in the sidebar. Each card shows the agents in the workflow, when it last ran, how many runs it has, and how it is triggered.

Building a workflow

Select Create Workflow to open the canvas. Add a step, then drag from the dot on its right edge to another step to say what runs after what.

Connections define order, not data plumbing. Anything a step depends on is passed to it automatically.

• Steps with no connection between them run at the same time.
• A step with two incoming connections waits for both to finish.
• Loops are rejected as you draw them, so a workflow cannot deadlock.

A workflow can hold up to 20 steps.

Configuring a step

Select a step to open its configuration.

Step name. How the step appears on the canvas and in run history.

Agent. Which agent runs this step. Different steps can use different agents, which is the point: a research agent, then a writing agent, then a messaging agent.

Prompt. What this step should do. The output of every upstream step is appended automatically, so you write the instruction for this step only and refer to the incoming work naturally, for example "summarise the findings above into five bullets."

Tools. Which of the agent's tools this step may use. Leave everything unchecked to allow all of them. Restricting tools is useful when a step should read but not send.

On failure. What happens when the step fails:

Setting	Behaviour
Fail the workflow	Stop the run
Retry, then fail the workflow	Retry, and stop if it still fails
Retry, then continue	Retry, and carry on either way
Continue	Move on without retrying

Success criteria. An optional plain-language description of what a good result looks like, for example "output has a price for each competitor." The step is checked against it before the workflow moves on.

Retries. Zero to three.

Max iterations. One to thirty. This caps how many tool-calling rounds the agent may take inside this step, so a step that gets stuck cannot run forever.

Running and scheduling

Open a workflow to run it, schedule it, and review past runs.

Run Now starts a run immediately. The graph updates as it goes, so you can watch which steps are queued, running, finished, or failed.

Three schedule modes are available:

• Manual. Runs only when you trigger it.
• Repeating. Runs on a recurring schedule.
• One-Time. Runs once at a time you pick.

Workflows can also be triggered through the API.

Run history

Every run is kept with its status, start time, duration, and how many steps completed. Open any run to see what each step produced, including runs from weeks ago. This matters when a scheduled workflow produces something unexpected and you need to find which step changed its behaviour.

Runs can be cancelled while in flight.

Workflow memory

When a step fails for a reason that is about behaviour, such as a wrong assumption, a missing instruction, or the wrong tool for the job, Arkios records what went wrong as a lesson and carries it into later runs of the same workflow.

Infrastructure failures are deliberately excluded. Timeouts, rate limits, and connection errors say nothing about how the agent should work, so they are never turned into lessons.

You can also add your own notes to workflow memory. A note is a standing instruction for every future run, which is the fastest way to correct something you keep seeing.

Notifications

Choose who hears about a run and when. Notifications can fire on every run or only when a condition you define is met. The person who created the workflow is always notified, and you can add other members.

A worked example

A weekly competitor brief, four steps:

1. Collect. A research agent with web tools gathers pricing and announcements from five competitors. Success criteria: output has a price for each competitor.
2. Analyse. A analyst agent reads the collected data and identifies what changed since last week. Depends on step 1.
3. Write. A writing agent turns the analysis into a one-page brief. Depends on step 2.
4. Send. A messaging agent emails the brief to the product team. Depends on step 3, tools restricted to Send Email, on failure set to retry then fail.

Set the schedule to repeat every Monday at 7am. Steps 1 through 3 can retry safely. Step 4 is the one that must not double-send, so it gets a single retry and stops the workflow if it still fails.

Related

• Jobs: single scheduled agent tasks
• Agents Overview: building the agents your steps use
• Tools and Connectors: giving agents access to your systems