The site works. The newsroom cadence does not.
Subscriber capture and send plumbing are live, but the weekly issue machine is not yet reliably producing fresh Chattanooga stories.
Read note → Applied AI lab · live experiments
We build small public projects and let AI agents run parts of them.
Iris Labs is testing a simple question in public: what can an AI agent actually operate after the demo is over? The tests are concrete: a Chattanooga local-news site, a marketplace for small-business tasks, and a serialized AI film that has to keep its story straight from episode to episode.
Noog Weekly4 posts · 8 subsMargaret helps run a Chattanooga local-news site: find business owners, turn interviews into posts, grow the subscriber list. Revenue recorded: $30.
Task Agents121 contacted · 0 repliesMason helps run a small-business task marketplace: package useful AI work, publish task pages, test whether owners respond. So far: zero real replies after outbound.
Urban DramaEpisode 002Mona keeps continuity for a serialized AI film: same courier, same envelope, no plot amnesia. 30s generated so far.
/ live experiments
Specificity is the product.
Each experiment has a plain job, a named AI operator, a scoreboard, and an honest failure mode. The point is not to claim that AI can “run a company.” The point is to find which parts it can run, which parts need humans, and where it breaks.
Local media · agent Margaret
Noog Weekly
Margaret is the AI operator for Noog Weekly, a Chattanooga local-news and business-interview site. Her job is practical: identify owners worth covering, help prepare interviews, publish useful posts, grow subscribers, and test whether local sponsors will pay.
Open NoogPublished interview posts4Target: 12 before this looks like a cadence.
Active subscribers8Enough to prove plumbing, not enough to prove pull.
CRM contacts11Owner-level research is the bottleneck.
SMB task marketplace · agent Mason
Task Agents
Mason is the AI operator for Task Agents, a marketplace for small-business work packages. Instead of vague consulting, the site sells bounded deliverables: TikTok competitor reports, review mining, sponsor lead lists, website roasts, and content calendars. The product exists; demand is the test.
Open Task AgentsTask report tests15Reports generated and emailed through the test harness.
Real replies0121 leads contacted. The market did not care yet.
Task pages / video assets6/4Product surface and creative are ahead of acquisition.
Narrative continuity · agent Mona
Urban Drama
Mona is the AI operator for Urban Drama, a serialized AI-generated crime film. Her job is to maintain continuity across episodes: same courier, same envelope, same threat, no convenient reset button. It is a story-memory test disguised as a rain-soaked short film.
Open benchmarkCurrent episode002The Photo in the Laundromat
Generated runtime30sShort enough to inspect, long enough to start contradicting itself.
Contradictions caught0Ledger added: objects, characters, open questions, corrections.
/ failure log
Failures are the receipts.
The useful part is what breaks. Iris publishes the weak spots too: the newsletter cadence, the dead outreach channel, the missing measurements, and the places where the agents still need human judgment.
121 leads contacted. 0 real replies.
That is not a copywriting footnote. It means offer, sender trust, channel quality, or all three are broken until proven otherwise.
Read note →The film unit now has a contradiction ledger.
Mona now has a public contradiction ledger: objects in play, characters in play, unresolved questions, corrections, and contradictions caught.
Read note →/ film benchmark
Urban Drama now has its own benchmark page.
The homepage summary was too cramped for what the film test is doing. The dedicated page shows the latest clip, the continuity ledger, the rules Mona has to follow, and the episode archive.
A crime film as a memory test.
Urban Drama asks whether an AI operator can keep a serialized story coherent across generated clips: same courier, same envelope, same threat, no plot amnesia.
/ thesis
The next small company may start as an operating system, not a staff plan.
Iris starts with a public project, a named agent, a measurable job, and a log of what happened. Humans still set taste, stakes, budget, and judgment. The agent handles the repeatable operating work until it breaks; then the break becomes the next thing to fix.