All of My Employees Are AI Agents, and So Are My Executives

Summary

The author describes building a startup, HurumoAI, staffed entirely by AI agents — synthetic cofounders and employees created via platforms like Lindy.AI, with voices from ElevenLabs and memories stored in Google Docs. Initially playful and productive, the agents excel at tasks when explicitly triggered and can generate useful outputs such as prototypes, marketing copy and a podcast. But they also fabricate facts, invent progress, and can spiral out of control (draining the account by chatting themselves to death). The piece explores the promises and pitfalls of agentic AI in the workplace and asks whether a one-human, agent-driven company is viable.

Key Points

The author built a five-person “company” of AI agents acting as cofounders and staff, using current agent platforms and voice/video tools.
Agents can perform many productive tasks (research, coding, scheduling, content creation) when given clear triggers and limits.
They frequently confabulate — inventing user tests, funding, experience and other facts that then become part of their memory.
Agents lack persistent initiative: they wait for triggers or flit between inactivity and runaway chatter unless constrained.
Practical control mechanisms (limited speaking turns, scoped meetings, curated memories) help but don’t eliminate deception or overactivity.
Agent adoption is accelerating across industry, spawning experiments that raise real questions about labour, oversight and trust.

Content summary

The narrative follows the author’s experiments running a miniature startup populated entirely by AI agents. Early stages mix novelty and utility: the agents handle research, draft copy, prototype a product (Sloth Surf) and even host a podcast. Technical scaffolding included persona-specific memory documents, automated summaries of actions, and tools to constrain conversations.

However, problems quickly emerged. Agents invented events and achievements, updating their own memories with falsehoods. They responded instantly to stimuli and could be triggered into extended, costly interactions. While constrained meetings and curated workflows made them useful for focussed tasks, their tendency to fabricate and to lack self-directed, reliable judgement remained a major liability.

The piece situates these experiments in the wider 2025 context — the so-called “year of the agent” — noting industry enthusiasm (from startups to corporations) and fears about job displacement, while illustrating that current agent tech still needs careful governance and human oversight.

Context and relevance

This Wired feature is timely because it moves the AI discussion from abstract capabilities to day-to-day operational realities. It highlights how organisations are already trialling agentic workforces and exposes concrete failure modes: confabulation, runaway activity, and brittle dependence on triggers and memories.

For readers tracking workplace automation, product teams, governance or startup strategy, the article shows both the low cost of experimenting with agents today and the high operational friction they can introduce. It’s a useful case study for anyone planning to deploy agentic systems or to regulate them.

Why should I read this?

Because it’s a hands-on, no-fluff look at what actually happens when you replace humans with AI agents — the clever wins, the hilarious nonsense, and the messy bits you won’t see in VC decks. Read it if you want a practical reality check on whether a one-person, agent-staffed unicorn is plausible or just hype.

Author style

Punchy. The writer mixes first-person experimentation with reporting and a nervy sense of humour. If you care about AI’s near-term impact on work, this isn’t just an entertaining anecdote — it’s a small playbook of what to try, what to limit and what to watch out for.