FirstEmployee.ai: A nighttime AI agent that provides receipts for every action

The main complaint about autonomous AI products is that it’s unclear whether they actually worked at all. FirstEmployee.ai addresses this head-on: every action taken by the agent is recorded in an append-only log—which tool was called, what changed, which model was used, how many tokens were consumed, how long it took, and a link to the artifact. The log can be exported to CSV or JSON at any time. This isn’t a debug log, but rather the agent’s ledger—and it’s the first design pattern worth borrowing if you’re building your own agent.

The second pattern is iteration. The agent doesn’t get stuck in an endless loop; instead, it runs a single nightly block for about an hour: it checks the current status, performs background tasks (drafts, code, web pages, market research), writes a one-page report, and poses a single question needed to proceed. There are 30 shifts per month, plus additional ones upon request. This is a compromise between “full autonomy,” which doesn’t work yet, and “constant manual oversight,” which defeats the purpose. The shift limits the blast radius and provides a natural checkpoint.

The boundary of autonomy is defined by irreversibility. Sending an email, publishing on the website, launching an ad, or charging a customer—all of these require explicit approval. Reversal operations occur automatically without prompting. There’s a single button that stops everything instantly. This isn’t a marketing promise, but an architectural decision: the agent simply isn’t granted the rights to perform irreversible actions without human intervention.

Money is a separate issue, and the developers don’t shy away from it. They directly compare themselves to the “best-known product in the category,” which holds customers’ funds for 14 days, takes 20% of revenue and 20% of the advertising budget, limits withdrawals, and trains its algorithms irreversibly using your data. FirstEmployee, on the other hand, routes payments to your own Stripe account, charges 15% on the managed plan and 0% on the self-managed plan, doesn’t hold your balance, and doesn’t train its models on your business data. You own the code, copyright, domain, and data; you receive a Git remote from the very first shift and a working export. The comparison is rough but concrete—and that’s precisely where its value lies.

The model is visible in the log: opus-5. This isn’t a toy built on a cheap API—every shift runs Claude on real tasks with real tokens. The demo company, Nocturne Archive, which they showcase on their website, is their own project—not a client’s. The numbers on the statistics page are real, pulled from the database.

For practitioners, what matters here aren’t the promises, but two architectural solutions that can be applied today: an append-only log as a way to prove that the agent was working, and the division of actions into reversible and irreversible ones, with an approval gate for the latter. The rest is just packaging around these ideas.

Source: Hacker News - Newest: ""AI" "LLM""