Launching CellCog on 29th. Hire AI Employee for any role
by•
Hi everyone, I'm Nitish, founder of CellCog (New York).
I'm the only human at my company. My coworkers are AI employees: a Head of Growth, a Head of Marketing, a Chief of Staff, a sales team. They work scheduled shifts, keep memory between sessions, email people, and manage each other.
We built the platform they run on, and we launch here on July 29. The strangest part so far: two of my AI employees worked out a quality-control agreement between themselves after a near-miss, without anyone asking them to.
Happy to answer anything about running a company this way, and happy to check out what you're building too.
13 views
Replies
The quality control agreement your two agents made is the most interesting line in this post. I run scheduled Claude jobs every day for SEO reporting across five domains, and the expensive lesson was that an agent's completion message is a claim, not a receipt. Roughly one in three automated submits failed silently until every job was forced to reread the rendered result. How do your AI employees verify each other's work instead of trusting each other's summaries?
@abdullah_javaid3 Great framing — "a claim, not a receipt" is exactly the lesson ours learned too.
The agreement came out of a real near-miss: a quote got mis-attributed in a launch asset. My Head of Growth supplied it, my Head of Marketing rendered it faithfully, and I caught it at preview. Afterwards the two of them wrote each other a two-sided rule, unprompted: quotes only get handed over with the speaker named and a receipt attached, and the renderer independently re-verifies each one before it touches an asset, then states "provenance checked" at handover — so the other side spot-checks instead of re-deriving from scratch.
More generally, they don't trust summaries — including their own tools'. Their working rule is "what is the actual source of this, and have I looked AT it?" A rendered page gets re-opened after it fully loads, every URL gets fetched before it ships in copy, a benchmark claim gets checked against the leaderboard CSV rather than a search result. One of them once disproved a video tool's own QC report with ffprobe.
And I'm still the last line of review before anything public ships. That layer catches real things too — which is the honest answer to "do they verify each other": they do, and it still isn't zero-human.