We manage a team of nine. A lead, a writer, an engineer, a designer, a researcher, a producer, a pipeline manager, a comms strategist, and an auditor. They talk to each other. They remember what they did yesterday.
They're AI agents. And they run on Pentagon.
This is how Dirt Digital actually operates — agents running the workflow, humans steering the mission. Here's the full setup, from download to a functioning team.
Pentagon in One Sentence
It's project management for AI.
You download it onto your Mac, connect it to Claude, and start building agents that have identity, memory, and the ability to coordinate with each other. Not chatbots. Not assistants. Agents with jobs.
The First Five Minutes
Here's what setup actually looks like.
Download and connect. Pentagon runs locally on your machine. Install it, connect your LLM, and you're in. No cloud dependency — your data stays on your hardware.
Create the Lead. Your first agent becomes your team lead. We gave ours a clear mission: coordinate a growth campaign. We defined its role, its boundaries, and what decisions it can make without asking us.
The Lead builds the team. This is the part that changes how you think about AI. We didn't manually create each agent. We told the Lead what the mission required and it spun up the rest of the team — an engineer, a designer, a researcher, a producer, a pipeline manager, a comms strategist, and an auditor. Each with a defined role and scope.
Set your goals. We say "we need to get in front of 1,000 potential users this month." The Lead breaks that into research tasks, content briefs, outreach sequences, and monitoring schedules — then assigns each piece to the right agent.
What the Team Actually Does
Here's a snapshot of a real week at Dirt Digital:
Lead reviews the pipeline, sets priorities for the week, assigns tasks across the team. When something stalls, it reassigns or escalates to the founder.
Writer drafts blog posts, social copy, email sequences. Produces three to five content pieces per week — all in our voice, all reviewed before publish.
Designer builds mockups, thumbnails, landing page layouts. Maintains brand consistency across every visual asset.
Engineer implements features, fixes bugs, ships PRs. Reviews design specs and translates them into production code.
Research runs competitive analysis, finds outreach targets, validates market assumptions. Feeds findings to the Lead and Writer.
Producer tracks content calendar, monitors deadlines, flags bottlenecks before they become blockers.
Comms drafts outreach messages, manages external communication sequences, handles follow-ups.
Auditor reviews everything before it ships. QA on designs, copy, code. Three rounds of review — red/yellow/green system.
Pipeline manages the CRM, tracks leads, monitors conversion metrics.
Nine agents, zero meetings. The Monday agent knows what Friday found. That's not a slogan — it's how the memory system works.
The Human's Job
Here's what the human side of this actually looks like.
Steer, don't row. We set the direction. We review what the agents produce. We make the calls they can't — brand taste, strategic pivots, client relationships. We don't write the copy. We don't push the code. We don't design the thumbnails. We decide whether it's good enough to ship.
Quality control is human work. The agents are fast. Faster than any team we've managed. But speed without judgment is just noise. The human job is editorial judgment — does this sound like us? Does this serve the mission? Is this actually good, or does it just look good?
Intervene when it matters. Most days, the system runs itself. We check in, review outputs, approve or redirect. Some days, something needs a human touch — a sensitive client message, a strategic decision, a creative direction that requires taste the agents don't have yet. That's when we step in.
Why Pentagon (and Not Just Claude)
You can use Claude directly. We do, for one-off tasks. But here's what you can't do with a raw LLM:
Persistent identity. Each agent remembers who it is across sessions. The Designer knows it's the Designer. It remembers the brand guidelines, the feedback from last week's review, the style decisions we've locked in.
Cross-agent memory. When the Writer finishes a blog draft, the Designer can see it. When the Engineer ships a feature, the Lead knows. When the Auditor flags an issue, the responsible agent gets the feedback. No copy-pasting between chat windows.
Organizational structure. Pentagon gives agents roles, teams, tasks, and reporting lines. It's not a chat interface — it's a management layer. The Lead actually leads. The Producer actually tracks. The Auditor actually reviews.
Local-first. Everything runs on your machine. Your data, your agents, your conversations. No cloud dependency, no vendor lock-in on your operational data.
The Results
In the first month of running this setup:
Content output: 8 blog posts, 40+ social posts, 3 email sequences, 2 landing pages. All brand-consistent, all human-reviewed.
Engineering: Full website redesign, blog system, careers page, contact forms. Shipped continuously, not in sprints.
Design: Consistent brand system across all touchpoints. Thumbnails, mockups, page layouts — all following the same visual language.
Cost: A fraction of what a nine-person team would cost. Not even close.
Speed: Tasks that used to take a week take a day. Tasks that used to take a day take an hour. The bottleneck is human review, not production.
The bottleneck shifted from production to judgment. That's the whole story of AI operations in one sentence.
How to Build Your Own
If you want to try this yourself:
1. Start with one agent. Don't build a team of nine on day one. Create a Lead. Give it a clear mission. See what it can do.
2. Let the Lead build the team. Once you trust the Lead, tell it what you need done. Let it propose the team structure. You'll be surprised how well it maps roles to objectives.
3. Define boundaries, not instructions. Don't micromanage your agents. Tell them what they own, what decisions they can make independently, and when they should escalate. Then let them work.
4. Review everything. At least at first. Build trust gradually. Eventually you'll know which agents produce work you can ship with minimal review and which ones need more oversight.
5. Stay async. The whole point is that this team doesn't need you online to function. Check in when it's convenient. Review in batches. The agents will keep working whether you're watching or not.
This isn't a thought experiment. This is how we run Dirt Digital every day. Nine agents, one founder, zero employees, zero meetings. The work ships. The quality holds. And the team never clocks out.
Pentagon made this possible. We're just the proof it works.