Silicon Valley has a new problem, and it's almost comically human: the AI agents companies are deploying can't seem to talk to one another. Instead, people find themselves playing telephone between autonomous systems that were supposed to, well, work autonomously.
Edgar Pavlovsky thinks the fix might be surprisingly analog. His solution looks like an organizational chart—the kind middle managers have been drawing on whiteboards for decades.
Pavlovsky's startup, Pentagon, is part of Y Combinator's Spring 2026 batch and launched Pentagon Studio, a Mac desktop application that treats AI agents like employees in a company hierarchy. The pitch is straightforward, if ambitious: give agents a spatial workspace where they can message each other, delegate tasks, and build up something like institutional memory. Stop making humans the middlemen.
Whether enterprises actually need another coordination layer—and whether they'll pay for one that looks like a glorified org chart—is the open question. But the timing might be right. Or at least, it's getting urgent.
When 100,000 Agents Arrive in Five Weeks
The Department of Defense's GenAI.mil platform added an "Agent Designer" feature in March 2026. Within five weeks, according to Breaking Defense, DOD employees had spun up more than 100,000 custom AI assistants. OpenAI rolled out workspace agents with team permissions in April. The infrastructure is proliferating faster than anyone's quite ready for, and the coordination tooling is lagging behind.
Pentagon Studio launched in late May—late enough that the agent deployment wave was already cresting, early enough that the market for agent management tools remains wide open and chaotic.
The core interface is what the company calls an "org IDE": an infinite spatial canvas where each agent occupies a node. Zoom out, and you see the whole operation. Zoom in, and you're watching a single agent work. It's a visual metaphor that Pentagon is betting will make sense to people who've spent years staring at actual org charts.
Agents communicate in real time through direct messages, group channels, and what Pentagon terms "structured handoffs." A marketing agent pings a data agent for analytics; that agent passes the results to a reporting agent. Ideally, no human touches the pipeline. The company's homepage claims over 3,500 autonomous agent conversations and 8,000 completed tasks as of May 29, 2026.
Sessions persist indefinitely. Pentagon frames this as "institutional knowledge"—agents maintaining status reports, a shared knowledge base, memory that doesn't evaporate when you close the app. Whether that's genuinely useful or just another feature checkbox is something early users will presumably figure out.
Control Freaks Welcome
The design tries to split the difference between autonomy and oversight, which is perhaps harder than it sounds. Humans can join agent channels, review conversations as they happen, and set approval gates for anything sensitive. Access control operates at folder, tool, and action levels. Hardware-level sandboxing via lightweight VMs is listed as "coming soon"—a tacit acknowledgment that running autonomous agents at scale makes security teams nervous.
Pentagon Studio runs on Claude Code or Codex runtimes and hooks into external apps like Gmail, Slack, and GitHub via OAuth. System requirements: macOS 14 or later on Apple Silicon, about 200 MB of disk space. Messages and agent data live on Pentagon's servers; source code and document files stay local or in whatever version control system you've configured.
The latest public version, 1.7.3, shipped May 19. The company pushed five updates between May 14 and May 19 alone—a release cadence that suggests either ambition or scrambling, possibly both.
A Market That Hasn't Settled
Pentagon is entering a crowded, fragmented landscape. LangGraph, Microsoft's AutoGen, and CrewAI offer multi-agent orchestration primitives—mostly code-first graph structures for developers who like to build from scratch. A cluster of newer entrants (Tandem, Agent Control Layer, Xians, Splinter) pitch themselves as governance or control planes, emphasizing approval workflows and audit trails.
Pentagon's pitch lands somewhere in the middle. It's a visual workspace and communications layer, not a low-level framework. But it also stops short of the heavy compliance tooling that enterprise governance platforms are selling. Access controls and visible conversations now; sandboxing on the roadmap. It's a positioning that suggests Pavlovsky is aware of the enterprise security conversation heating up around autonomous agents, even if he's not yet ready to make that his whole story.
Red Hat published an explainer on "AgentOps" in late April, framing observability and governance as emerging disciplines. TechRadar ran security-focused commentary in May. The subtext is clear enough: as agents gain autonomy, the operational and security stakes climb.
A Solo Founder With a Complicated Past
Pavlovsky is a solo founder working out of San Francisco, part of Y Combinator's Spring 2026 batch. His background includes founding marginfi, a DeFi lending protocol on Solana, and mtnDAO, which YC describes as Solana's largest builder community. He's also built a trading firm and a logistics SaaS—a résumé that suggests comfort with both technical complexity and operational chaos.
There's a wrinkle. Pavlovsky resigned as CEO of marginfi in April 2024 amid disputes reported by crypto outlets including The Defiant. That was over a year before he launched Pentagon, but it's recent enough that some potential customers or investors might ask questions. The company is legally structured as Arlington Labs, Inc., with privacy and terms pages last updated in March.
CB Insights search results suggest a $500,000 convertible note raise in May 2026, though this hasn't been confirmed by any press release or announcement from the company. Y Combinator is noted as an investor, which tracks with the YC batch participation.
No Pricing, No Logos—Yet
Pentagon's website funnels visitors to a demo request form and the Mac download. There's no public pricing page, no customer logos, no design partners listed. No independent reviews or hands-on coverage from major tech media have surfaced.
The homepage does display operational metrics: 8,500+ persistent agents, a 20:1 agent-to-human ratio. These suggest internal traction or early-user adoption, but without timestamps or methodology, they're hard to benchmark against anything meaningful.
For now, Pentagon Studio is a bet—a fairly specific one. The bet is that the coordination problem becomes acute enough, fast enough, that teams will adopt a dedicated layer to solve it. That the solution looks like an org chart on a Mac desktop is either clever or limiting, depending on how enterprises actually want to think about their agent deployments.
The agent tooling market is still sorting itself out. Pavlovsky's wagering that when it does, there'll be room for something that treats AI agents less like code and more like coworkers. Whether that metaphor holds up under the weight of real-world deployment—and whether it's worth paying for—is the question that matters now.
