Source: raw/How_to_Build_an_AI_Agent_Team_for_Automated_Workflows_AI_Agents_Tutorial.md (auto-caption transcript)

Creator: AI Master (YouTube @iamAImaster) · URL: https://www.youtube.com/watch?v=bJ70Grz2ydM · Published: not in the transcript · Sponsor: HyperAgent, the platform the build runs on. Confidence is low because this is a sponsored single demo. The video does not say which model powers the agents.

AI Master builds a three-stage YouTube research chain on HyperAgent, a hosted multi-agent platform. Renee collects AI release news each weekday morning, Miles verifies and compares it, and Casey proposes three content directions. Each agent has a named job, a plain-English job description, its own cloud login, and Slack as the handoff surface. The design choices carry over to other platforms: a strict job description, a reviewed manual run before scheduling, corrections saved as skills, a defined output shape, and a final human approval gate.

Key Takeaways

  • One stage per agent, and the next agent’s input is the previous agent’s output. Renee’s posted briefing is Miles’s input, and Miles’s page is Casey’s. “Nobody copies anything into a new window, and I don’t re-explain the context.” Miles “has a different job, not just a longer version of Renee’s.”
  • The job description is “where all the leverage sits.” “Monitor AI news and summarize it” gives “something technically correct and completely useless.” Renee’s paragraph names a source list (AI model release pages, developer changelogs, two trusted newsletters). It limits scope to meaningful changes in features, pricing or availability. It fixes a per-item format: date, source link, one line on what changed, one line on why it matters to the channel.
  • Schedule it, but run it by hand first. Renee runs every weekday morning before work, in the cloud on her own login, posting to a #ai-desk Slack channel. Before leaving it unattended, the creator ran her once and checked each item against the job description. He opened source links to compare her summary with the release page.
  • Correct in the channel, and the correction becomes a skill. He replied to Renee in Slack explaining why an item missed the criteria and asked her to keep the rule. She saved it as a skill that carries into future runs, so the job description sharpens without being rewritten. He still reads every run.
  • Define the final output’s shape before the agent runs. The creator calls this “the single most useful decision in this whole build.” Miles verifies Renee’s flagged releases against primary sources and compares competing models on pricing, context limits and chosen benchmarks. His output is a page: one card per model with key numbers, source links next to the numbers, and a verdict with reasoning at the bottom. He tags the creator in Slack. “If your agent’s answer is a paragraph in a chat window, you’ve moved the work rather than removed it.”
  • The last agent stops before publishing. Casey turns Miles’s page into three directions in the same format: concept, two lines of coverage, core angle, why it matters to the viewer, rough section flow, and links back to specific findings. No script and no upload package. “Anything an agent would send or publish waits on a yes from a human,” set as a permission when creating the agent. “If I never answer, nothing goes out.”
  • Proactive agents are possible. A second HyperAgent example, Vera, a data analyst in Slack, noticed an activation problem unprompted. She posted a short message with an attached analysis report containing charts and recommended actions.
  • Agents can be shared. Per the video, HyperAgent agents can be shared with collaborators so they use the same context, skills and memory.
  • Naming is not cosmetic. “The moment a job has a name attached to it, you stop thinking about prompts. You start thinking instead about who owns what.”

Try It

  1. Pick one repeatable job. Define where it starts and what a finished result looks like. Build one agent around that.
  2. Write the job description as a contract: sources, inclusion rule, exclusion rule, per-item format, schedule and delivery channel.
  3. Run it once manually and audit it item by item against the description before you schedule it.
  4. Add a second agent only when a stage depends on the first one’s result. The same shape works outside content, for example incoming data → assessment → client report.
  5. Specify output as a reviewable artifact (cards, sources beside numbers, verdict last) and a delivery location.
  6. Gate anything that sends or publishes behind your explicit approval.

Implementation

Tool/Service: HyperAgent (hosted multi-agent platform). Agents run in the cloud, each with its own login, and connect to Slack, Google Docs and Gmail.

Setup: Create an agent → name it → write the job description → set the schedule → connect Slack and choose a channel → set the publish/send permission.

Cost: Per the video, signing up for a paid plan through its link gives $100 in bonus credits. Plan prices are not stated.

Integration notes: The Slack channel serves as both the handoff bus between agents and the review surface for the human.

Open Questions

  • Which model or models run HyperAgent agents, and whether you can choose.
  • How skills saved from Slack corrections are stored and whether they can be edited directly.
  • Pricing beyond the bonus-credit offer.