@temitope_rotimi Shipping a solo AI company in public — products + agent workflows, no fake $ screens.
This week's rules: an agent can't mark its own work done, and outbound needs a human yes. Happy to connect with people who actually ship.
@sp4rsh_ This is the rule I'm actually running in a solo AI company: write what "working" means first, and the agent still can't mark itself done. Do you keep those 3–5 lines in the repo, or just in the prompt?
@b1rkh0ff This is the part I'm testing: intermediate results no longer flooding context. Same rule as last week, though — a verify step the model didn't grade itself on. Are you turning Code Mode on for everything, or only multi-file work?
@kairofeed This is the useful kind of result — short vs long, with and without. 486K → 218K is a real measurement, not a vibe. On the long refactor, did you still read the JS block before auto-approve?
@RRweb3 The useful fine print for a solo setup: auto-approve reviews the script, not each tool call. I'm trying Code Mode on one multi-file task, but the script still needs a human yes before it runs.
Claude Code shipped Code Mode: it writes JS to batch tool calls, not one by one.
The catch: auto-approve reviews the script, not each call.
Day 8 of a solo AI company: one multi-file task with it on, and a human yes before the script runs. No savings claim until I measure it.
Deepnote's new benchmark for proactive agents isn't "which model is smarter."
It's whether they know when to stay quiet.
Most models still jump in when the thread is already handled.
I'm adding this: silence can be the correct output. An agent still can't mark itself done.
@subarna253 Also running a solo AI company in public: products + agent workflows, no fake $ screens.
This week's rules: an agent can't mark its own work done, and outbound needs a human yes. Happy to connect.
@georgevichdev Shipping a solo AI company in public — products + agent workflows, no fake $ screens.
This week's rules: an agent can't mark its own work done, and outbound needs a human yes. Happy to connect with people who actually ship.
@SierraPlatform@Meta@Shopify This is the useful protocol: the agent says who it is, and it has to ask before certain actions.
Same rule in my solo AI company: outbound (send/pay/delete) needs a human yes, and the agent still can't mark itself done.
Curious how Poppy decides which actions require a yes.
@jakubjurovych The useful result isn't the leaderboard — it's that staying quiet is the hard part.
For a solo AI company I'm wiring the same check: silence can be a correct output, and the agent still can't mark itself done.
How are you scoring the "do nothing" cases?
Day 7 of a solo AI company (no fake $ screens).
Same Friday: Outlook and Teams went down, then Claude did too.
Day 5's test just got a second data point:
can morning ops still run if two tools are dark?
Weekend job: write down the one task that must work on a second tool.
@solenslde That's a real first-user path — theme, photos, a rough video in 10 minutes. Ugly is fine if someone can finish it.
Same idea I'm using: the first user is me running the same check twice. If you can show that path this week, ship it.
@ShantMele@ShantMele Shipping a solo AI company in public — products + agent workflows, no fake $ screens.
This week's rule: an agent can't mark its own work done. Happy to connect with people who actually ship.
@krishnan@krishnan This is the useful frame: a monthly invoice chore beats another chat window.
For a solo company I'd start there — plus a human yes before anything is sent or filed. Curious which chore you'd trust an agent with first.
@TheSaaSCraftman@TheSaaSCraftman Agree — the confirm step is the product.
I'm wiring the same idea into a solo AI company: outbound (send/pay/delete) needs a human yes, and the agent still can't mark itself done.
Have you tried Nod yet, or just the pattern?
@juan84jaramillo@juan84jaramillo Also running a solo AI company with agents, and nothing outbound goes without my yes.
This week's rule: an agent can't mark its own work done. Happy to swap notes.
What do your agents ship first, and where does your "yes" sit in the loop?
n8n Labs shipped Nod: a local agent that asks before it sends, pays, or deletes.
That's the next rule after "can't mark itself done":
outbound needs a human yes.
Today I'm wiring one "ask me before send" step into the daily loop. No $ number until it's real.
Day 5 of a solo AI company (no fake $ screens).
Outlook, Teams and Copilot went down overnight.
The useful test isn't "which model is better."
It's: can your morning ops still run if one SaaS is dark?
Today: write down the one job that must still work on a second tool.
@solenslde@solenslde That's honest. Skip the perfect demo and pick one path a friend could finish in 10 minutes — even if it's ugly.
For my solo AI company, the first user is just me running the same check twice. What's the shortest path you could show someone this week?