
Agents for Humans:
Building Citereel, an Evidence-Backed Product Video Agent on AWS
How I built an agentic video-production workflow for the Agents for Humans hackathon with Strands Agents, Amazon Bedrock AgentCore, and a queue-driven AWS rendering pipeline.
A product can have a thoughtful website, a clear value proposition, and a real story to tell - and still never make it into a video. For a founder or small product team, video production usually means a long handoff chain: write a brief, find the right pages, draft a storyboard, capture the product, record narration, edit the footage, and check every claim. The work is slow, expensive, and easy to get wrong when the copy is generated from assumptions instead of the product itself.
That is the problem I set out to solve for the Agents for Humans hackathon.
I built Citereel, an evidence-backed video-production application for founders, marketers, and product teams. A creator starts with a product URL. Citereel inspects the authorized public site, proposes a storyboard, shows the source material behind each claim, and waits for the creator to review it. Only after approval does it capture real browser footage, create narration, and render a product video.
The goal is not to make an agent that takes over creative judgment. The goal is to remove repetitive production work while keeping people in control of the story they publish.
The design decision that shaped the project
The first question was not "How do I generate a video?" It was "Where should an agent be allowed to make decisions?"
Citereel uses Strands Agents for the reasoning-heavy part of the workflow: turning approved product context into a grounded plan. The agent organizes source material, suggests scenes, drafts narration, and explains why a scene belongs in the story. The creator can inspect the proposed claims and storyboard before recording begins.
That review step is deliberate. A generated video can look polished while still saying something incomplete, outdated, or wrong. Citereel treats the agent's plan as a proposal, not as permission to publish. The human approves the material, and the application records the result of that decision alongside the production.
This created a useful product boundary: the agent plans; the creator approves; the system executes and reports what happened.
The AWS workflow behind Citereel
Citereel is deployed as a production-style AWS workflow rather than a single long-running web request.
At the front of the system, Amazon CloudFront serves the Citereel Studio from Amazon S3. The studio calls an API built with Amazon API Gateway and AWS Lambda. Amazon DynamoDB stores production state, approvals, claims, and the job record.
Once a creator approves a storyboard, the application puts rendering work onto Amazon SQS. This queue separates the interactive experience from the slower browser-capture and media-rendering steps. A dispatcher coordinates work, and an Amazon ECS on AWS Fargate worker runs the isolated browser capture and video-processing tasks. Rendered files and supporting artifacts are stored in S3.
The agent itself runs through Amazon Bedrock AgentCore Runtime. AgentCore Runtime is designed to host agent applications and supports frameworks including Strands Agents. It also provides isolated session environments, which is useful when an application needs to keep one creator's planning context separate from another's. Amazon Bedrock AgentCore Runtime documentation
The architecture is intentionally split by responsibility:
- Strands Agents and AgentCore Runtime handle planning and orchestration.
- API Gateway, Lambda, and DynamoDB manage the interactive application state.
- SQS and Fargate handle durable, asynchronous production work.
- S3 and CloudFront deliver the studio and finished assets.
- Amazon Polly provides narration in the production pipeline.
This separation made the project easier to reason about. A slow render should not lock up the creator's browser. A browser-capture task should not share the same lifecycle as a request to review a storyboard. And an agent's proposed copy should remain visible and reviewable before it becomes a finished video.
Why Strands Agents fit this product
The hard part of Citereel is not calling a model once. It is maintaining a workflow with clear checkpoints:
- Inspect approved product context.
- Build a grounded storyboard and draft narration.
- Show sources and claims to the creator.
- Wait for review and approval.
- Capture the authorized pages in an isolated browser.
- Add narration, render the video, and run a final quality check.
Strands Agents gave me an agentic layer for planning without forcing the production system into a free-form chat experience. The agent can reason about source material and form a coherent sequence of scenes, while the rest of the application remains explicit, observable, and easy to recover when a task fails.
Hosting that agent on AgentCore Runtime also made the deployment model clearer. AgentCore Runtime is framework-agnostic and supports code-based agents built with Strands. It provides a managed runtime for the agent while the rest of Citereel continues to use the AWS services that fit each production responsibility. AgentCore Runtime overview
What changed during the build
The project became better when I stopped treating the video as the only output.
Early on, it was tempting to focus on the final render: a nice looking clip is easy to understand in a demo. But the more important product behavior is the path to that clip. Citereel surfaces the storyboard, source evidence, approvals, job state, and a report of the resulting production. That makes the workflow reviewable rather than magical.
I also learned to make the agent's scope narrower. Asking an agent to "make a great product video" is vague. Asking it to propose a sequence from authorized pages, attach the relevant source context, and wait for approval creates a system that is more predictable and easier for a creator to trust.
The queue-based worker design was another important shift. Browser capture and media rendering take time, and they can fail for ordinary reasons such as a page loading slowly or a worker needing to be retried. Moving that work to SQS and Fargate gave the application a clean handoff between the creative plan and the production machinery.
What I would build next
The next version of Citereel would deepen the creator controls rather than simply add more generation. I would add stronger scene-level editing, reusable brand kits, clearer revision comparisons, and more production analytics. I would also keep improving the evidence trail so a creator can answer a simple question at any point: "Where did this line in my video come from?"
That question is central to the project. AI can speed up creative work, but it should leave people with more clarity and agency.
Citereel is my entry for the Agents for Humans hackathon: a system where a Strands Agent helps turn a product's public story into a video, while AWS provides the durable, scalable pieces that make the workflow real.
Project: Citereel
Public code: https://github.com/vvdotcom/Citereel
Demo video: https://youtu.be/kb56yzrBgxk
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article