Workflow automation
Automations you can read afterwards.
Workflows are described in plain English and then exist as inspectable steps. A dry run tells you what it would do and what it would cost before anything happens, and every real run keeps per-step forensics.
- 1 Find the contact and their last message lookup
- 2 Decide whether it still needs an answer classify: resolved / waiting on us / waiting on them llm.judge
- 3 Draft the follow-up llm.generate
Estimated cost $0.41 for 37 conversations · nothing has run
Or just ask for it
Describe the automation. Read it before it runs.
A dry run and a cost, in the thread where you asked.
InboxDry run: 47 conversations, est. $1.12
When a conversation goes quiet for two days, classify why and either nudge them or close it.
Built it as six steps you can read — idle trigger, AI classify, branch, nudge, close, log. Dry run over the last week: 47 conversations, 31 nudged, 16 closed, est. $1.12. Nothing was sent. Save it?
@Octo add an approval step before it emails anyone
On it 🔧(runningworkflow_dry_run…)
Added before step 5. It now stops and waits for a person, and I re-ran the dry run — same 47, no sends. (edited)
🔒 34 replies · last reply 2m ago
good. route the approvals to this channel
Email octo@agent.omniocto.com · WhatsApp 1-MAN-ASK-OCTO · Slack #omniocto
- Describe-it-in-English builder
- Say what should happen and the workflow is built from it, then shown to you as steps you can read rather than as a black box.
- Eleven trigger kinds
- Schedules, webhooks, idle conversations, labels, tags, every inbound message and more — so a workflow can start from something happening rather than from a person remembering.
- Over sixty step kinds
- The building blocks for real automations rather than a handful of toy actions.
- AI steps
- Classify, judge, extract, summarise and generate as ordinary steps, so a workflow can make a judgement in the middle without leaving the workflow.
- Human-approval steps
- A workflow can stop and wait for a person, which is what makes it safe to automate something consequential.
- Transfer-to-human
- Hand the conversation to a person as a step, on the call or in the thread rather than in an email an hour later.
- Dry run with cost estimate
- See what it would do and what it would cost before it does it — the difference between confidence and hope on a workflow that messages customers.
- Per-step run forensics
- Every run records what each step received and produced, so a wrong result is diagnosable rather than mysterious.
- Surgical edits
- Change one field without regenerating the whole workflow, so a small correction does not risk everything else in it.
- Slack actions
- Post and act in Slack as workflow steps.
- HTTP, GraphQL and signed-webhook calls
- Call your own systems from a workflow, with signed webhooks where the receiver needs to verify the sender.
- OAuth & API-key connections
- Stored credentials for the services a workflow reaches, rather than keys pasted into steps.
- Sub-workflows
- Compose workflows out of other workflows instead of copying steps between them.
What this powers