Multi-Agent Workflows

​

Overview

• Composable steps: Each step belongs to a team, has a type, and exposes statuses (with one default).
• Semantic automation: Status transitions can be triggered by inputs; completed statuses emit outputs. You wire outputs → inputs to automate orchestration across steps.
• Multi-agent execution: Invoke Tessact’s first-party agents (e.g., remix, compliance, captioning, thumbnail generation) or bring your own third-party AI via connectors and webhooks.
• Safety by design: Orchestration enforces acyclic flows (no infinite loops) and validates wiring before publish.

​

Core Concepts

​

Steps, Types, and Statuses

• Step — A unit of work owned by a team (e.g., Ingest, Edit, Compliance Review, Publish).
• Type — Optimised configuration for that step’s work mode. Examples include:
  • Watch Folders — Monitor Connected Folders and react to file events.
  • AI Agent — Run one or more agents (e.g., Generate YouTube thumbnails, Blog post from webinar, Compliance scan).
  • Video Creation — Editor-centric step for timeline edits and remixes.
  • Manual Review — Gate with assignments, SLAs, and decision outcomes.
  • Publishing — Prepare deliverables and push to destinations/CMS.
  • Utility / Transform — Transcode, proxy, sprite sheet, metadata extraction.
  • Webhook / External Action — Call external systems and wait for callbacks.
• Statuses — Each step has a default status (e.g., Not started), plus additional statuses (e.g., In progress, Needs edit, Approved, Rejected). Statuses power both visibility and automation.

​

Inputs & Outputs (Wiring)

• Inputs are triggers that cause a step’s status to change automatically (e.g., AgentCompleted, FileMoved, ReviewApproved, TimeElapsed, WebhookReceived, MetadataUpdated).
• Outputs are signals emitted when a status is reached (e.g., Edit.Completed → outputs.Done).

• Connect Output → Input to propagate progress across steps.
• Multiple Outputs → one Input act as an AND-gate: the Input fires only when all connected Outputs are true.
• One Output → multiple Inputs (fan-out) is allowed.
• No loops: A step status cannot simultaneously use its own IN and OUT as triggers. The builder validates flows as a DAG (Directed Acyclic Graph).

​

First-Party and Third-Party AI Agents

​

Designing Workflows

​

Typical End-to-End Pattern

1. Ingest (Watch Folders / Library) → asset appears and analyses automatically.
2. AI Agent Pack → captioning, compliance, thumbnails, social remixes.
3. Manual Review → approval gate with assignments and SLAs.
4. Publish → deliver to CMS/social, archive to Connected Folder.

​

Advanced Patterns

• Editorial factory: AI Search for Clips → Remix → QC → Publish.
• Knowledge expansion: Blog Generation from Video → SEO title & description → Manual copy edit → Publish.
• Compliance with regional overrides: Compliance EU AND Compliance US → both must pass before Publish.

​

Create a Workflow (Step-by-Step)

​

Attach to a Project

• Create or open a Project → Workflow → select the published workflow.
• Provide any step-level parameters required by chosen types (e.g., which Connected Folder to watch, or agent presets).
• On save, the project enters the workflow’s initial status and begins execution when triggers fire.

​

Manual Review Layers

• Assignments — Round-robin, owner-based, or group-based.
• Checklists — Define criteria (e.g., brand safety, loudness).
• SLA timers — Overdue states emit alerts and can auto-escalate.
• Decisions — Approved, Needs edit, Rejected emit Outputs you can wire forward.

​

Execution & Monitoring

• Run History — Timeline of step/status changes, agent jobs, and callbacks.
• Retries & Backoff — Transient failures auto-retry with exponential backoff; hard failures surface with actionable logs.
• Idempotency — External calls include a run-scoped idempotency key to avoid duplicate work.
• Alerts & Webhooks — Subscribe to events (e.g., Step.Failed, Publish.Completed).

{
  "event": "workflow.step.failed",
  "workflow_id": "wf_3bpv9",
  "project_id": "proj_9f2mx",
  "step": "Compliance Review",
  "status": "Failed",
  "attempt": 3,
  "run_id": "run_1a2b3c",
  "timestamp": "2025-08-25T10:45:12Z"
}

​

Performance & Cost Management

• Concurrency — Limit concurrent agent executions per workflow or step.
• GPU quotas — Reserve or cap GPU-bound steps for predictable throughput.
• Caching — Reuse deterministic outputs (e.g., transcripts, embeddings) across runs.
• Agent-compute minutes — Usage accrues per agent runtime; see org-level analytics.

​

Next Steps

• Create a Workflow
• Attach to a Project
• Manual Review