Agentic Workflows

1. General

Proposed By: Shane Burgett

Proposed On: 2026-05-28

2. What Is The Problem?

One specific sentence. Not a theme — the actual friction or gap.

Agents do not have a simple, reusable way to create and run Livepeer-powered video and media workflows without building a dedicated application for each task.

3. Why Does It Matter To the Livepeer Ecosystem?

Several sentences on why solving this problem matters, including one sentence on the cost of not solving it or the current window.

Livepeer Modules make it possible to expose paid media compute, GPU capacity, runners, discovery, and billing as reusable network capabilities, but those capabilities are still too low-level for agents to compose directly inside their workflows. Today, an agent that wants to do something practical like slide extraction on a live stream, meeting capture and summarization, or home-video indexing must stitch together ingest, storage, model execution, workflow logic, payment, events, artifacts, and search as a custom integration. If this gap remains, Livepeer risks being powerful infrastructure that agents cannot easily use at the moment when agentic workflows are becoming a primary interface for software automation.

4. What Does Success Look Like?

Concrete and observable. "X goes from Y to Z" or "teams can now do X without Y." No aspirational language. Use bullet points for multiple outcomes.

• An agent can start a video or media workflow with one stable interface, such as input + workflow + session, instead of manually integrating Livepeer APIs, runners, payment, and storage.

• MVP workflow packs that enable some primary capabilies like live-stream slide extraction, meeting capture and summarization, realtime security alerts, or video search can be launched from reusable workflow packs rather than bespoke applications.

• Livepeer Modules remain the underlying compute, media, and billing layer, while agents interact with higher-level sessions, events, artifacts, and search results.

• Developers can add new model or media capabilities as Livepeer modules or workflow blocks without changing the agent-facing UX.

• The setup path for an agent goes from “build a custom media application” to “install a skill/CLI, choose a workflow, customize it provide an input, and follow results.”

5. Open Questions

Unknowns, assumptions to test, or decisions still needed before this can move forward. One question per bullet.

• What is the smallest agent-facing interface that can cover live streams, uploaded assets, local files, browser capture, and model workflows without exposing Livepeer internals?

• What workflow representation should agents use to describe media jobs in a portable way, regardless of whether the execution engine is Roboflow, a custom Livepeer workflow runner, or another tool?

• Which compute-heavy model capabilities should Livepeer expose first so agents do not need to assemble their own GPU stack, such as object detection, OCR, transcription, embeddings, segmentation, or YOLO-style runners?

• Should workflow execution happen as one paid Livepeer session, as separate paid block/model calls, or as a hybrid of both?

• What parts can be built as a thin external layer today, and what would require new Livepeer modules, runner contracts, SDK changes, or workflow-engine integrations?

• How should agents receive durable outputs such as clips, transcripts, summaries, detections, embeddings, alerts, and searchable timelines?

• What default workflows should prove the concept first: security stream analysis, meeting capture and memory, home-video search, slide extraction, or another high-signal use case?

• What security, privacy, and consent requirements apply when agents capture browser tabs, meetings, cameras, or personal video libraries?

• How much should the agent own locally versus delegate to hosted or networked Livepeer infrastructure?

• What minimum developer experience would let an agent use Livepeer for a new media automation without creating a new vertical application or custom integration each time?