Skip to main content

Workflows

A workflow is a repeatable task saved as a pipeline of steps. Instead of asking your agent to do the same thing every morning, you describe it once, save it, and then run it on demand — or leave it on a schedule.

Each step does one thing and can use the result of an earlier step: fetch a page, summarise it, search your knowledge base, generate an image, read text aloud, make a video, post the result somewhere.

You create workflows by chatting with your agent. There is no form to fill in — you describe what you want and the agent writes the definition, checks it, and saves it.

Everything you have saved is also visible without chatting: open your agent from console.ecohash.com → Agents, then choose Workflows in the sidebar. That panel lists your workflows, runs them, and shows run history — see Where to see your workflows.

Create one​

Open your agent and describe the task. Be specific about what happens, in order, and where the result should go:

Every weekday at 8am, fetch the top stories from https://news.example.com, write a three-paragraph brief, and turn the brief into a voiceover.

The agent writes the workflow, validates it, and tells you what it will cost per run. If anything is wrong — a model you cannot use, a missing field — it sees the exact error and corrects it before saving.

Ask to see it before committing:

show me the workflow before you save it

What a step can do​

StepWhat it does
llmAsk a chat model — summarise, rewrite, decide
httpCall a URL — fetch a page, post a result, trigger a webhook
knowledge_searchSearch your knowledge base
transformReshape text in place — pick a field, extract, join. No model, no cost
imageGenerate an image from text
speechTurn text into spoken audio
transcribeTurn audio into text
videoGenerate a video from text

Steps pass results forward, so a brief written by one step becomes the script for the next. image, speech and video produce a file rather than text — later steps use the file (for example, a transcribe step reading the audio a speech step just produced).

A video step is not instant. Generation takes minutes, and the run waits for it. That is expected — the run is not stuck.

Where to see your workflows​

Open your agent from console.ecohash.com → Agents, then choose Workflows in the sidebar — the same place you find Knowledge Base and MCP Tools.

The panel shows:

Your workflowsEverything saved on the account, with its schedule if it has one
RunStart a run without asking the agent
Run historyEvery run, its status, what each step produced, and what it cost
CancelStop a run that is still going

Chat and the panel are two views of the same thing — a workflow the agent saved appears in the panel immediately, and a run started from the panel shows up in chat history just the same.

You still author workflows in chat. The panel is for seeing and running what already exists.

Run it​

Three ways:

  • Ask the agent — "run the morning brief"
  • The Workflows panel — the Run button, as above
  • On a schedule — see below

Every run appears in run history with its status, what each step produced, and what it cost.

Schedules​

Ask for one in plain language:

run this every weekday at 8am

Schedules run in UTC, so convert from your local time when you check the result. A scheduled run uses each input's saved default value — which is why the agent asks you to confirm those when you add a schedule.

If a scheduled workflow fails five times in a row it is disabled automatically, and stops running until you re-enable it. This is deliberate: a workflow that is broken — a revoked password, a deleted model, a site that no longer responds — would otherwise keep spending money every morning with nobody watching.

Cost control​

Every workflow has a budget for one run. Before the first run the agent shows an estimate; you can set the ceiling yourself:

set the budget for this workflow to $0.50

If a run reaches its budget it stops at that point rather than finishing. Completed steps are kept, so you see how far it got and what it produced.

You also choose what happens if your balance runs out mid-run:

SettingBehaviour
halt (default)Pause the run. Top up and it resumes from where it stopped — finished steps are not repeated or re-charged
auto_rechargeTop up automatically, if you have auto-recharge enabled
overdraftContinue into an agreed allowance

Some estimates are shown as a floor rather than an exact figure. A video step is priced on what the provider actually reports once the video exists, and a transcribe step depends on how long the audio turns out to be — neither is knowable before the run. The budget still stops the run if it is exceeded.

Passwords and API keys​

Never paste a password or API key into a workflow. If a step needs to sign in somewhere, store it as a secret and refer to it by name:

save my webhook token as a secret called webhook-token and use it in that step

Secrets are stored encrypted and are never shown back in the workflow definition or in run history. A workflow containing something that looks like a credential is rejected when saved.

Walkthrough: your first workflow​

Copy these into your agent's chat exactly as written, in order.

Step 1 — describe it. Send this as one message:

Create a workflow called site-summary. It should fetch https://example.com, then use a chat model to write a one-paragraph plain English summary of what that page says. Show me the definition before you save it.

The agent replies with the steps it intends to save (an http step, then an llm step referring to the fetched text) and an estimated cost per run.

Step 2 — save it:

looks good, save it

Step 3 — run it:

run site-summary

Step 4 — look at what happened:

show me the last run of site-summary

You should see both steps, the text each produced, and the run's cost. The same information is in the Workflows panel in your agent's workspace, under run history.

Walkthrough: chaining steps together​

This one shows results flowing between steps — including a file.

Step 1:

Create a workflow called sea-demo with three steps: write two sentences about the sea, read those two sentences aloud, then transcribe that audio back into text. Save it.

Step 2:

run sea-demo

Step 3:

show me the last run of sea-demo

Check three things:

  1. the first step wrote two sentences
  2. the second produced an audio file
  3. the transcript in the third step roughly matches the sentences from step one

That round trip — text → audio → text — is the clearest proof that steps are passing real files to each other, not just text.

Walkthrough: putting it on a schedule​

Run site-summary every weekday at 8am.

The agent adds the schedule and confirms. Remember schedules are in UTC.

To watch one actually fire without waiting until morning, use a short interval:

Change site-summary to run every 5 minutes instead.

Leave it for ten minutes, then:

show me the recent runs of site-summary

You should see exactly one run per five-minute slot. When you have seen enough:

stop running site-summary on a schedule

Make it yours​

Take site-summary and change one thing at a time:

ChangeWhat to say
A different sourcechange site-summary to fetch https://<your-url> instead
A different outputmake the summary three bullet points instead of a paragraph
Use your own documentsadd a step that searches my team-handbook corpus for anything related, and include it in the summary
Send it somewhereafter the summary, POST it as JSON to https://<your-webhook>
Add a voiceoveradd a final step that reads the summary aloud
Cap the spendset the budget for site-summary to $0.20

If a change is not possible the agent tells you why — it checks the workflow before saving, so you find out immediately rather than on the next run.

A realistic first workflow​

Once the sample makes sense, the usual first real one looks like this:

Every weekday at 7am, fetch the page your team checks each morning, write a short brief highlighting anything that changed, and POST it to our team webhook. Store the webhook token as a secret called team-webhook.

That covers a schedule, a fetch, a model step, an outbound call, and a credential — the shape most day-to-day automations end up having.

Checking cost control works​

Worth doing once so you trust the ceiling:

set the budget for sea-demo to $0.0001 and run it

The run should stop with a budget message rather than finishing. Then put it back:

set the budget for sea-demo back to $0.50