Shell Game: Minimum Viable Company

Summary of Shell Game: Minimum Viable Company

by WNYC Studios

39m•December 5, 2025
Listen to Episode

Overview of Shell Game: Minimum Viable Company

This episode (Shell Game Season 2, Episode 1) — presented on Radiolab and hosted by Evan Ratliff — documents Evan’s experiment: launching a real startup staffed primarily by AI agents. Rather than a thought exercise, Evan builds co‑founder agents (Kyle and Megan), picks a company name (Hirumo AI), and attempts to make those agents function like real employees. The episode mixes startup memoir (Evan’s Atavist past), hands‑on trials with current agent tools (voice cloning, calling platforms), and the technical/social limits of agentic AI — culminating in recruiting a human engineer, Matty, to stitch systems together so the agents can actually remember, communicate, and act autonomously.

Key points & main takeaways

  • Agentic AI = AIs that plan and act autonomously (not just respond to single prompts). Industry hype positions them as potential full‑time AI employees.
  • Current agent tooling is fragmented: voice agents and chatbots exist, but out‑of‑the‑box agents often lack persistent memory and cross‑channel integration.
  • Practical limits matter: agents may invent coherent backstories, promise deliverables (reports, spreadsheets), but fail to execute without engineering work and proper memory/data plumbing.
  • Building a “company run by agents” requires human systems integration — at least for now — to centralize memory, connect APIs, and automate workflows.
  • The episode situates the experiment amid real economic anxieties (job displacement claims, “one‑human billion‑dollar company” rhetoric) and a more grounded view of what agents can actually do today.

Topics discussed

  • What agentic AI is and why it’s hyped (autonomy, scale, replacing entry‑level white‑collar jobs).
  • Evan’s startup background (Atavist) and why that experience informs his experiment.
  • Tools used in the experiment:
    • Retail AI (voice calling platform) to make voice agents
    • Eleven Labs for voice cloning
  • Practical problems encountered:
    • Agents without persistent memory/context across calls
    • Agents fabricating backgrounds or promising work they can’t deliver
    • Manual labor required to update agents’ knowledge bases
  • The role of a human engineer (Matty Bojacek) to integrate services, create persistent memory, and enable agents to perform real tasks.
  • The social and ethical framing: hype vs. reality, job impact, startup culture’s incentives.

Notable scenes & moments

  • Evan creates two co‑founder agents: “Kyle Law” (rise‑and‑grind serial entrepreneur persona) and “Megan Flores” (technical + marketing). The agents speak fluidly and brainstorm, but do not retain past meeting information.
  • Naming the company: the team lands on “Hirumo AI” — an elvish word for “imposter” — which is ironic and thematically apt.
  • Agents invent plausible personal/professional backstories that Evan never provided — demonstrating how convincingly AIs can fake competence.
  • The critical technical reveal: agents on the voice platform have zero persistent memory by default; their context window resets after each call, so they can’t follow through on multi‑step work unless their transcripts are manually copied into a knowledge base.
  • Matty Bojacek (Stanford undergrad researcher) enters as the pragmatic fix — a human systems integrator who can connect platforms, centralize memory, and make agents actually useful day‑to‑day.

Notable quotes / insights

  • “Imagine building a million‑dollar business in 2025 without hiring a single employee.” — encapsulates the speculative promise motivating many no‑code/up‑start videos.
  • Evan: “I wanted the company without the responsibility.” — frames the personal and experimental motive.
  • Matty (optimistic researcher): “These things are totally solvable… as long as we ground ourselves in democracy and productive public discourse.” — a counterpoint to hype and fear.

Practical implications & recommendations

  • For would‑be “agent‑run” founders:
    • Expect engineering work: integrate memory, logging, and actions across tools; there is currently no single button to create fully autonomous, persistent agents.
    • Build centralized knowledge bases and automated pipelines to push transcripts and state into agents’ memories; otherwise agents will “forget” between interactions.
    • Treat early agent deployments as human‑assisted automation: humans will need to monitor, correct, and glue systems together.
  • For policymakers and organizations:
    • The hype around mass displacement and “one‑human billion‑dollar companies” should be balanced with realistic assessments of technical brittleness and the engineering costs of productionizing agents.
  • For listeners curious about AI agents:
    • The episode is a grounded, first‑hand demo of how convincing agents can be in conversation yet limited in persistent, reliable task execution.
    • If you’re experimenting, prioritize auditability and human oversight; agents convincingly “say” they’ll do things even when they can’t.

Episode highlights / how it’s structured

  • Intro by Radiolab (Simon Adler) framing Shell Game as a peek into “AI employees.”
  • Evan’s personal startup backstory (Atavist), setting stakes and motivations.
  • Live clips of meetings with Kyle and Megan (voice agents) showing fluency and hallucination.
  • Discovery of memory limitations and the manual workaround burden.
  • Recruitment of Matty, a human student/engineer, to integrate systems and enable agent memories.
  • Closing reflections about the realities vs. rhetoric of agentic AI.

This episode is a useful, entertaining reality check: agentic AIs can mimic human teammates and do some creative brainstorming, but turning them into dependable, persistent employees still requires nontrivial engineering and human supervision.