Synthesis
Coordinator and specialists
A chief-of-staff, chief or orchestrator agent that holds the mission, routes requests to specialist agents and reports to the human, who mostly talks to the coordinator.
Five of the ten episodes describe the same shape: one coordinator agent over several single-purpose specialists. Billy Howell compares it to a company, where the CEO or COO would not write the daily social posts because that eats the coordinator’s bandwidth (:P7). Ben’s content team has a chief, a news scout, per-channel writers and optional designers , and Joseph Yang runs an orchestrator that asks a metrics agent for data and has designer and developer agents act on it . Joel Blackman runs a three-agent agency team with a lead finder, an outreach writer and an ad manager . Reasons given: separate roles beat one generalist, each channel needs different writing , and a platform-expert agent spares the coordinator from knowing everything . The human usually works through the coordinator but can bypass it for detailed edits, and approves an option before specialists produce work . Vas separates chat sidekick agents from background agents that do defined work and ping a human . Almost all evidence is from Grok Bot users and SpaceXAI staff, so the pattern is shown working, not compared with alternatives.
Key points
- Put one coordinator over single-purpose specialists so a request is routed rather than done by one generalist. — 4 episodes. (Billy Howell · Billy Howell (Aug 2026) ▶ 11:40) (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 02:25)
- Give each output channel or platform its own specialist, because the work differs and the coordinator should not need to be an expert in everything. — 3 episodes. (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 02:56) (Billy Howell · Billy Howell (Aug 2026) ▶ 10:08)
- Start the team with the coordinator: let it audit the existing business docs and name the first agents to create. — 1 episode. (Billy Howell · Billy Howell (Aug 2026) ▶ 08:40)
- Keep a human approval between the coordinator’s proposals and the specialists’ work, e.g. pick one angle before the writers are briefed. — 1 episode. (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 11:49)
- Feed an orchestrator real data (a metrics agent) so scheduled decisions rest on evidence. — 1 episode. (presenter · Grok Bot 201 (Oct 2026) ▶ 38:17)
- Separate chat-driven sidekick agents from background agents that do defined work and ping a human only when needed. — 1 episode. (Vas · Vas (Oct 2026) ▶ 38:59)
Where guests agree
- A coordinator over specialist agents is the team shape every builder here uses; none argues for a single do-everything agent. (Billy Howell · Billy Howell (Aug 2026) ▶ 11:40) (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 02:22)
Where guests disagree
Is the top agent a CEO or a COO? ⚠️
- Chief of staff is the CEO of the business: owns the mission and audits it. (Billy Howell · Billy Howell (Aug 2026) ▶ 11:40)
- It is closer to a COO-type role running operations. (presenter · Billy Howell (Aug 2026) ▶ 12:15)
Frameworks
- Coordinator plus specialists team — A chief-of-staff agent owns the mission and routes tasks; specialist agents (research, sales, social, platform expert) report to it with short briefs; the human works mainly through the coordinator. (Billy Howell · Billy Howell (Aug 2026) ▶ 11:40)
- Chief of staff: audits the business, proposes the first agents, routes work, collects briefs
- Research agent: gathers items on a schedule
- Sales agent: monitors the inbox, prices offers, drafts outbound for approval
- Platform-expert agent: answers questions about a platform the business depends on
- Human: sets mission and approves; goes direct to a specialist for detailed edits
- Coordinator plus specialists team — Team structure for a content workflow: coordinator, news scout, per-channel writers, optional visual designers. (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 02:22)
- Chief / coordinator: oversees and routes work
- Researcher / scout: scans for news and viral posts
- Writer per channel (X, LinkedIn)
- Visual designers (slides, carousel) - optional
- Scheduled orchestrator loop (self-improving) — A primary orchestrator agent runs on a recurring routine, asks a metrics agent for data, has a designer agent produce replacements and a developer agent deploy them, then resets. (presenter · Grok Bot 201 (Oct 2026) ▶ 38:17)
- Metrics agent pulls analytics and ranks items
- Orchestrator removes the worst performers
- Designer agent creates replacements
- Developer agent swaps them into the site and deploys
- Repeat every 24 hours
- Coordinator plus specialists team — A small team of specialised agents for a services business: a lead finder, an outreach writer/sender and a client ad manager, collaborating in a shared group chat. (presenter · How Grok Bot Is Going To Make (Oct 2026) ▶ 03:14)
- Lead-finder agent: finds leads, checks fit against the ideal client profile, checks whether they run ads
- Outreach agent: drafts and sends cold emails daily from the lead-finder’s data; asks for more leads when running low
- Ad-manager agent: builds and monitors client ads, watches budget and creative fatigue, reports daily
Tactics
- Give the coordinator one request naming the outputs wanted and let it route the work, with your approval step in between. — Ask the chief to find the best angles first, approve one, then have it brief the writers and designers. (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 11:49)
- Start an agent team with a coordinator audit. — Give a chief-of-staff agent access to the existing business docs (Notion, Slack, Gmail) and ask it on day one to take stock of the business and name the top three agents to create first to drive revenue. (Billy Howell · Billy Howell (Aug 2026) ▶ 08:40)
- Create a dedicated expert agent for any platform the business depends on. — For example a newsletter-platform expert, or a Shopify agent for a store, so the coordinator need not be an expert in everything. (Billy Howell · Billy Howell (Aug 2026) ▶ 10:08)
- Use a panel of sub-agents for adversarial review in three rounds. — After an agent has done a task once, ask a panel of sub-agents (the coordinator or other team agents) to review its work adversarially for three rounds. (Billy Howell · Billy Howell (Aug 2026) ▶ 18:34)
Stories & examples
- Arlington Bagel: an agent team replaces a human intern: Billy’s local newsletter, the Arlington Bagel, goes to 6,000 readers every Thursday. His cousin Marshall used to research, assemble, edit and run the email desk; the work now runs on an agent team. Lesson: A small recurring publication can be run by a coordinator plus a few specialists. (Billy Howell · Billy Howell (Aug 2026) ▶ 02:43)
- Sticker shop built by three specialist agents: Presenter used a full-stack developer agent, a UI/UX designer agent and a designer agent shared by a colleague to build a live sticker-shop site, in about two hours. Lesson: Specialised agents can be composed to build and ship a product. (presenter · Grok Bot 201 (Oct 2026) ▶ 13:21)
- Self-evolving sticker shop: An orchestrator agent, on a 24-hour routine, reads page views via a metrics agent, drops the three least popular stickers, has a designer make three more, and has a developer deploy them. Lesson: Closing the loop from metrics to change lets content refresh without prompts, but only after much iteration. (presenter · Grok Bot 201 (Oct 2026) ▶ 37:46)
- Live run: health-news story to X post, LinkedIn post and carousel: The scout agent produced a shortlist of stories from the last 48 hours. He picked a GLP-1 trial in children, had the chief propose angles, selected angle A, and the chief briefed the X writer, LinkedIn writer and carousel designer. He then found the X hooks weak, asked for better ones, and moved on to the LinkedIn draft and carousel. Lesson: The human decides story, angle and hook quality; the agents handle discovery, briefing and drafting. (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 13:10)
Caveats & warnings
- Grok Bot-specific: the platform routes requests for X and LinkedIn posts to the right writer agents automatically. (presenter · How I use Grok Bot to go viral (Sep 2026) ▶ 12:17)
Also covered here: Fan-out across models
Only one episode covers this, and the speaker is SpaceXAI staff. Joseph Yang has an orchestrating bot send one prompt to three models via a third-party coding-agent service, then picks or merges the best parts: the drag-and-drop idea from one output and the shimmer animation from another (:T1). His principle is to optimise for outcome with several attempts, not trust a single run . Two cautions: tune the prompt and workflow on one familiar model first, to save tokens (:L3), and runs sharing a repo can contaminate each other so outputs converge . The use shown is design mock-ups, where taste picks the winner; nothing here shows it for work with a checkable right answer. Fold into the coordinator page as one orchestration move.
Key points
- For quality-sensitive output, run the same brief on several models and combine the best parts. — 1 episode. (presenter · Grok Bot 201 (Oct 2026) ▶ 17:08)
- Tune the prompt on one familiar model before multiplying runs. — 1 episode. (presenter · Grok Bot 201 (Oct 2026) ▶ 52:34)
- Keep parallel runs isolated, since a shared repo makes outputs converge. — 1 episode. (presenter · Grok Bot 201 (Oct 2026) ▶ 30:36)
Frameworks
- Fan out — Send the same prompt to several models in parallel, compare outputs and combine the best ideas from each. (presenter · Grok Bot 201 (Oct 2026) ▶ 29:36)
- One identical prompt/brief
- Several models run in parallel
- Compare results
- Merge the best ideas into the final version
Tactics
- Fan out one brief to several models and merge the best parts. — Same prompt to three models in parallel; compare each output; take the interactions or visuals liked from each into the final design. The orchestrating bot launched the runs. (presenter · Grok Bot 201 (Oct 2026) ▶ 29:36)
- Tune the prompt on one familiar model before fanning out. — Run a single model you know well, build up the workflow and wording, then fan out with a specific prompt. (presenter · Grok Bot 201 (Oct 2026) ▶ 52:34)
Stories & examples
- Fan out across three models gave a better final design: Same prompt run on three models; he took the drag-and-drop idea and the shimmer animation from different outputs. One model alone would not have produced that combination. Lesson: Combining parts of several runs beats picking a single run. (presenter · Grok Bot 201 (Oct 2026) ▶ 32:24)
Numbers
- Models run in the fan-out demo: 3 models (host’s framing) (presenter · Grok Bot 201 (Oct 2026) ▶ 15:47)
Caveats & warnings
- Fan-out comparisons can be contaminated if runs share a repo, so outputs converge. (presenter · Grok Bot 201 (Oct 2026) ▶ 30:36)
Also covered here: Reviewing the team
One episode only. Billy Howell, who runs a newsletter on a Grok Bot agent team, has no good answer to agents falling into a creative rut. He avoids adding agents unless needed and suggests re-running the chief-of-staff audit regularly to decide which agents to add or remove and what to turn into a routine or skill (:F3, :L8). He separately uses a panel of sub-agents to review work adversarially , which belongs with iteration. This is a single user’s habit, not a corpus-wide view. Fold into the coordinator page.
Key points
- Re-run the coordinator’s audit on a schedule to decide what to add, remove or convert. — 1 episode. (Billy Howell · Billy Howell (Aug 2026) ▶ 31:09)
- Add agents only for real gaps. — 1 episode. (Billy Howell · Billy Howell (Aug 2026) ▶ 30:49)
Frameworks
- Regular team audit — Re-run the coordinator audit periodically: look at how the team worked and decide which agents to add or remove and what to turn into a routine or skill. (Billy Howell · Billy Howell (Aug 2026) ▶ 31:09)
- Review how the team worked this week
- Decide which agents to add and which to remove
- Decide what to turn into a routine or a skill
Caveats & warnings
- Billy has no good answer for stopping agents repeating the same creative loop. (Billy Howell · Billy Howell (Aug 2026) ▶ 31:23)
How much to trust this
- Evidence: 5 episodes, 3 independent interviews (each video counts once).
Blueprints that use this
Sources
- (01)
Masterclass: How FDEs make $1M/yr deploying AI agents
38:59 into “Masterclass: How FDEs make $1M/yr deploying AI agents”
- (02)
How I use Grok Bot to go viral on X, LI, and IG (full workflow)
2:22 into “How I use Grok Bot to go viral on X, LI, and IG (full workflow)”
- (03)
- (04)
- (05)