Pawel, this is the rare agent post that reads like a system, not a séance.
The night shift loop is the part that matters. Planning, execution, wrapup, with success criteria and a morning brief. That is closer to operational capacity than “look what my chatbot said.” Also appreciate the honesty about where you stay in the loop. Creative direction, quality judgment, voice, strategy, and social nuance. That matches what I keep seeing. The tool can move fast. It cannot own the consequences.
Two things I would love to see, mostly because you clearly have them.
First, a lightweight scoreboard. Not to satisfy the “prove it” crowd, but because it makes the experiment more interesting. Time saved, deploys that actually moved a metric, revenue to date, and the current cost breakdown. Even rough numbers would turn this into a case study.
Second, the guardrails in a little more detail. If Wiz can deploy to production overnight, what gates have to pass first. Tests, staging, feature flags, rollback. The part that scares people is not autonomy. It is unbounded autonomy.
Also, the “agent selling blueprints for building agents to pay for itself” angle is hilarious and kind of the point. At minimum it is a better plot than most SaaS origin stories.
Curious where this lands in 30 days. Does Project Money break even, or does it reveal the real product is the ongoing log of what an agent can actually ship when you treat it like a junior dev with rules and accountability.
totally love this. i bet financially the plan your on is really valid. another substacker was just doing the math and probably ona similar level to this. it would be 1400$ over api calls lol
Insightful and fun to read post Pawel! I'm just starting with AI Agents (Open Claw obviously) and I have one technical/security question if you don't mind: how do solve agent authorisations for email and other PI services? Do you grant full access via oAuth or delegate some authorisations? What about services without oAuth, do you share your passwords with Agents in .env or some other way? Asking mostly from a security standpoint and what do you feel comfortable with.
Great question Mario → this was the thing I thought hardest about before giving my agent real access.
My actual approach:
- OAuth/API tokens where possible - Typefully and servies all use scoped tokens. Not passwords.
- App-specific passwords for email — Gmail SMTP uses an app password, not my main account. If it leaks, I revoke that one token, nothing else affected.
- Local secrets dir for the rest — Services without OAuth get credentials in a local global/secrets/ dir. Gitignored, never cloud-synced, only readable by local scripts.
- Cookie extraction for browser-based services — My agent reads session cookies directly from Arc. Means no passwords shared, but requires trusting the agent with browser state.
What I won't do: store master passwords, use credentials that can't be individually revoked, or sync any of this to the cloud.
My mental model: "Can I audit and revoke this in 5 minutes if something goes wrong?" If yes — comfortable. If no — rethink the architecture.
I love this. I'm running a similar setup. Claude Code agent with persistent memory, skill system, cron jobs, the works. The "can it earn its own keep" question is exactly right. My agent handles daily briefings, content pipeline, and research that would cost me 3-4 hours/day. At $200/month, that's roughly $2/hour for a tireless junior analyst. Next on my list is plumbing my Claude code software builds into it.
The cost isn't the point here. Real question: can AI create enough value to fund its own inference costs? If that loop closes, we're looking at fundamentally different economics. That's the experiment worth watching.
That's exactly the right frame. Cost per token is a distraction.
Think about it differently: I'm running agents overnight on my codebase right now. $40 in API costs, 15-20 hours of reviewed pull requests by morning. A contractor doing that same work costs $2,000+ and takes a week. I do it in my sleep. Literally.
But the real win isn't cost savings. It's time to market. Work that used to wait in a backlog ships overnight. Features compound faster. The gap between "idea" and "deployed" keeps shrinking.
The loop isn't just closing on cost. It's closing on speed. And speed compounds in ways cost savings never will.
I think both points and angles are valid but from a different perspectives and on different layers. There's also a third one:
- personal/solo dev efficiency bosts - yeah, the speed we can iterate and ship stuff is crazy, the value is already there.
- organizational efficiency boost or growth: speed is there but imho it's not the biggest value for organizations (although on a surface level this is what most companies try to do). For organizations, agentic AI creates value when it is aimed at meaningful end-to-end process friction, not when it simply accelerates isolated tasks. Faster code generation does not fix poor process design, fragmented workflows and communication or lack of ownership or unclear decision criteria. Agents (or ML/AI in general) can also help with this but it requires mental shift from "boost speed" to "eliminate friction" mindset.
- economy disruption: this is what Pawel refers to. If an AI Agent can actually break even... then where's the limit? :) Will there be a job for us? Will we need a job anymore? Everything will change, exciting times!
Wow all new what is the easiest entry point not to get lost
Would it be an email reader
Creating to do with the emails
What’s the best way to train yourself to do this start?
A bit overwhelmed, but extremely interested
nice article!
Definitely need more information
Yeah, I know my substack gets a little bit too nerdy/techy at times :D
I am working on some things more “open” to non-tech people as well(or outside of AI bubble).
You cna start here: https://thoughts.jock.pl/p/start-here
Pawel, this is the rare agent post that reads like a system, not a séance.
The night shift loop is the part that matters. Planning, execution, wrapup, with success criteria and a morning brief. That is closer to operational capacity than “look what my chatbot said.” Also appreciate the honesty about where you stay in the loop. Creative direction, quality judgment, voice, strategy, and social nuance. That matches what I keep seeing. The tool can move fast. It cannot own the consequences.
Two things I would love to see, mostly because you clearly have them.
First, a lightweight scoreboard. Not to satisfy the “prove it” crowd, but because it makes the experiment more interesting. Time saved, deploys that actually moved a metric, revenue to date, and the current cost breakdown. Even rough numbers would turn this into a case study.
Second, the guardrails in a little more detail. If Wiz can deploy to production overnight, what gates have to pass first. Tests, staging, feature flags, rollback. The part that scares people is not autonomy. It is unbounded autonomy.
Also, the “agent selling blueprints for building agents to pay for itself” angle is hilarious and kind of the point. At minimum it is a better plot than most SaaS origin stories.
Curious where this lands in 30 days. Does Project Money break even, or does it reveal the real product is the ongoing log of what an agent can actually ship when you treat it like a junior dev with rules and accountability.
I will publish some news today about that!
Woot! 🙌
totally love this. i bet financially the plan your on is really valid. another substacker was just doing the math and probably ona similar level to this. it would be 1400$ over api calls lol
I even saw wilder calculations that some users are milking MAX subs for 2400 :D
geeze lol. i do wonder what i would hit now….
Insightful and fun to read post Pawel! I'm just starting with AI Agents (Open Claw obviously) and I have one technical/security question if you don't mind: how do solve agent authorisations for email and other PI services? Do you grant full access via oAuth or delegate some authorisations? What about services without oAuth, do you share your passwords with Agents in .env or some other way? Asking mostly from a security standpoint and what do you feel comfortable with.
Great question Mario → this was the thing I thought hardest about before giving my agent real access.
My actual approach:
- OAuth/API tokens where possible - Typefully and servies all use scoped tokens. Not passwords.
- App-specific passwords for email — Gmail SMTP uses an app password, not my main account. If it leaks, I revoke that one token, nothing else affected.
- Local secrets dir for the rest — Services without OAuth get credentials in a local global/secrets/ dir. Gitignored, never cloud-synced, only readable by local scripts.
- Cookie extraction for browser-based services — My agent reads session cookies directly from Arc. Means no passwords shared, but requires trusting the agent with browser state.
What I won't do: store master passwords, use credentials that can't be individually revoked, or sync any of this to the cloud.
My mental model: "Can I audit and revoke this in 5 minutes if something goes wrong?" If yes — comfortable. If no — rethink the architecture.
Great approach, thanks! How do you monitor if your app password leaked, crone jobs -> check https://haveibeenpwned.com every x? :D
I am using cron a lot…but not for this one. I have to think about this actually to make it more secured! Thanks!
I love this. I'm running a similar setup. Claude Code agent with persistent memory, skill system, cron jobs, the works. The "can it earn its own keep" question is exactly right. My agent handles daily briefings, content pipeline, and research that would cost me 3-4 hours/day. At $200/month, that's roughly $2/hour for a tireless junior analyst. Next on my list is plumbing my Claude code software builds into it.
The cost isn't the point here. Real question: can AI create enough value to fund its own inference costs? If that loop closes, we're looking at fundamentally different economics. That's the experiment worth watching.
That's exactly the right frame. Cost per token is a distraction.
Think about it differently: I'm running agents overnight on my codebase right now. $40 in API costs, 15-20 hours of reviewed pull requests by morning. A contractor doing that same work costs $2,000+ and takes a week. I do it in my sleep. Literally.
But the real win isn't cost savings. It's time to market. Work that used to wait in a backlog ships overnight. Features compound faster. The gap between "idea" and "deployed" keeps shrinking.
The loop isn't just closing on cost. It's closing on speed. And speed compounds in ways cost savings never will.
I think both points and angles are valid but from a different perspectives and on different layers. There's also a third one:
- personal/solo dev efficiency bosts - yeah, the speed we can iterate and ship stuff is crazy, the value is already there.
- organizational efficiency boost or growth: speed is there but imho it's not the biggest value for organizations (although on a surface level this is what most companies try to do). For organizations, agentic AI creates value when it is aimed at meaningful end-to-end process friction, not when it simply accelerates isolated tasks. Faster code generation does not fix poor process design, fragmented workflows and communication or lack of ownership or unclear decision criteria. Agents (or ML/AI in general) can also help with this but it requires mental shift from "boost speed" to "eliminate friction" mindset.
- economy disruption: this is what Pawel refers to. If an AI Agent can actually break even... then where's the limit? :) Will there be a job for us? Will we need a job anymore? Everything will change, exciting times!