Workflow bot demos

Run your first demo

1. Choose a workflow such as Deployment incident, Sales follow-up, or Billing operations.
2. Name your bot, select a provider and model ID, and adjust its instructions and workspace lifetime.
3. Choose Launch workflow. Blobfish assembles a real seeded Twin environment under your identity.
4. Use a saved provider connection, an operator-configured connection, or enter a provider key for this browser session. An unavailable model returns an explicit error.
5. Send a starter prompt or your own message. Watch tool calls in Chat or Activity. Open Workspace to inspect the state the bot changed.
6. Send follow-up messages, reload, or return to the bot from the sidebar. The same conversation and workspace continue.

Browser demos use the existing Twin limits: two live environments, three model turns per hour, and six tool steps per turn. Account and API-key identities use the configured step limit, up to 20. Each turn is bounded to three minutes.

When a chat reaches its tool limit, the bot gets one final response with tools disabled to summarize what it observed and identify unfinished work. Blobfish displays the limit notice and records skipped tool requests separately from executed actions. This summary remains ungraded; it does not certify that the workflow succeeded.

Available presets

Choose from 21 bot setups. The catalog covers e-commerce checkout, SaaS billing and upgrades, DevOps CI/CD, customer support, document collaboration, payment fraud, access reviews, high-volume messaging, internationalization, and scheduling conflicts, plus Salesforce follow-up. Search by workflow, scenario name, app, or tag.

Variants add abandoned checkouts, duplicate credits, an already-tracked release, approval boundaries, finance sharing, partial refunds, contractor offboarding, an alert flood, conflicting refund announcements, and a Berlin–Tokyo handoff. Each starts with different records and includes editable bot instructions and starter prompts.

Preset availability depends on mounted twins. Coverage notes are shown before launch. The demo uses the existing Twin runtime and does not claim to have provisioned a separate cloud container.

Steer, stop, reset, and replay

During an active turn, send a clarification. The UI distinguishes a queued instruction from the bot’s delivery acknowledgment. Stop cancels the running turn; incomplete writes are discarded before the next turn. In Setup, Reset opens a new conversation at the original scenario baseline. Old conversations and their execution tapes remain available from Conversation history.

Activity exports a Blobfish JSON document containing bot configuration, conversation messages, tool events, completed-turn identifiers, and workspace state. Provider keys and private thinking blocks are excluded. Tool actions can be replayed; receipts that omit private model content are labeled and cannot reproduce the exact model request. Free-form chat has no task verifier, so it is shown as ungraded. Use the environment’s task runner for measured rewards.

API

All bot operations are owner-scoped. Use your existing Blobfish API key or the browser session cookie. Saved bot configuration accepts no provider secrets.

POST /api/v1/twin/bots
{
  "template_id": "deployment-incident",
  "name": "Release engineer",
  "provider": "openai",
  "model": "gpt-5-mini",
  "max_steps": 12,
  "ttl_minutes": 60
}

To attach to an existing owned Twin environment with the required apps, include env_id. Each bot gets a separate conversation session.

POST /api/v1/twin/bots/{bot_id}/chat
{
  "conversation_id": "<from bot.conversations>",
  "turn_id": "turn_<unique_request_id>",
  "message": "Investigate the failed check on PR #482."
}

Chat streams start, step, steer_ack, complete, and error events. A repeated completed turn ID replays its receipt. A different prompt with that ID is rejected. Optional api_key and base_url are request-scoped; custom web endpoints require public HTTPS.

Paths in the table are relative to /api/v1/twin. Manage environment lifetime and termination from Twin environments.