2025 will be the year we see the first self-driving startups.
Level 0: No AI
People do everything. They come up with ideas, build products, and run operations. Many legacy businesses still work this way.
Level 1: People use AI tools ⬅︎ we are here
People might use ChatGPT to help write copy or Cursor to help write code. This is where most startups are today.
Level 2: AI agents complete tasks based on human instructions
People might ask AI agents to write software from a plain-English spec or tell it execute well-defined customer service processes. At this point entire departments (like support or QA) get largely replaced by AI. No startups I know of operate at this level yet—but if yours does, let me know.
Level 3: AI agents propose changes to their own instructions
They might propose new customer service processes and product changes in response to customer feedback. Humans would still approve each of those changes. Just a few people could run a large company this way.
Level 4: AI agents autonomously change their instructions
At this point startups become self-improving. Humans would only be involved as an escalation point or where required by the real world (e.g. to raise capital or to incorporate). At this point many startups would only have one human.
Level 5: No humans
AI agents decide which businesses to start, raise capital (through crypto tokens or other means), build and run them. No humans required. This would require major reforms in the legal and financial system.
@dinasaur_404 is there any way we could lift the restriction of having only five cloudflare accounts per user?
our preferred deployment topology for iterate is to make a brand new cloudflare account for each deployment
our installation wizard helps people do that by creating a cloudflare account for them
unfortunately we cannot test this reliably, because our cloudflare users already have five accounts and it doesn't seem possible to delete them
@EastlondonDev@jhutchings0@dinasaur_404@justbobsledit Yeah that is using the old tenant API. I might try that. There's a few new options (account creation API and even "temporary account for AI agents"), but they don't work very well at all yet :(
@jhutchings0 any chance I could get 15 mins of your time? I think we're using a bunch of these APIs how you'd like them to be used, but encountering a fair bit of friction
Our idea is broadly that
- we want to make self-improving digital minds
- they should each be confined in an isolated cloudflare account
- the self-improving mind should have access to almost every part the cloudflare account (except a very sensitive secrets vault worker)
- the account should get its money from a human owner
- we should encourage all users of our software to use this deployment topology
I hope you are planning to have more open cloudflare accounts than there are humans on the planet!
btw we also have some feedback on the account creation API and related flows
1. in order to create an account for a user, we need a cloudflare access token. to obtain it, we do a cloudflare oauth consent asking just for their user info. but during that flow cloudflare asks users to select an account, even though no account scopes were requested. this is confusing
2. after creating a new account, the screen for adding payment methods often shows an error for a couple of minutes (something something eventual consistency)
3. similarly, after adding a $5/month workers subscription, it sometimes takes a while until a worker can be deployed
4. we can't use the "temporary agents accounts" abstraction, because our software requires dynamic workers, containers, r2 and a dedicated domain. all of those are paid features
5. IDEALLY we'd like to just set up the whole account for our users, complete with paid workers plan, some starting credit for ai gateway loaded etc and THEN transfer the account to the user. But there is no API for creating accounts with a credit balance at scale
I now use my @Pebble ring ~50 times a day
Goes to my personal iterate mind, which uses Jev to decide whether to forward on to my work mind or keep it
My personal and work mind exchanged a keypair with each other and forwarded events are signed
It works insanely well and took 5 mins to set up. I just asked "Can you somehow exchange keys with the mind at hostname X and make it so work related pebble recordings go to that mind"
Now I just walk around and can say "We are out of bin bags" immediately followed by "Signup should just be an MCP server" and it all just happens. Bin bags will arrive and signup will be an MCP server
We have a "coding chief of staff" agent context in our company that is kinda like our merge queue / CI
It tells other agents when they can merge what in what order and what they need to improve
Sometimes we would talk to specific low level agents directly (because it's lower latency)
So it started subscribing to receive copies of those other agents input events immediately, just so it is never in the dark, and then immediately follow each message we send to a subagent with their own follow up message
E.g. after I said "Yes, do that" to an agent, the chief of staff immediately followed up "I saw Jonas's word to you: he APPROVES the Cloudflare TURN relay"
We're re-inventing middle management from first principles
A big fintech irony in Europe: Shitty banks that let you approve MasterCard 3DS / visa secure txs via email or SMS code suddenly have the best user experience for agents
I can give my agent an email address and phone number, but can't do fancy in-app approvals
Of course it is quite insecure, but sadly there doesn't seem to be a better way to say "I am willing to entrust this agent with £50 to buy my groceries and if they lose the money, so be it"
The @browsercompany has been prototyping at the intersection of (a) full-duplex voice models & (b) traditional software GUIs. “Chat” just feels too limiting for most tasks.
We stumbled upon something interesting & wanted you to be able to play with it. Demo below (audio on):
@NathanFlurry@rivet_dev Oh cool! I asked my computer to try running iterate on rivet on Friday and it said it had to make loads of synchronous functions async. Should have double checked
It would be really cool if Cloudflare shipped controlplane env bindings
Imagine deploying a worker with an https://t.co/PU3khbdMSu binding that lets you hit the controller plane without secrets
Permissions would have to be subset of the key that was used to deploy the worker
Then you could run an AI agent in a cloudflare account and truly allow it to self-improve - without secrets
@KentonVarda wdyt?
A couple of other things that that would help agents with nearly total control of a cloudflare account to self-improve safely
1) oauth / api keys that are authorised to “edit any workers except except X”
2) a cf-owned soft hsm / secrets management service that can
- generate secrets and key pairs with no codepath to
- do cryptographic operations using the keys
- fetch _through_ the credential to a list of approved hosts
So even owners of the cloudflare account should only be able to put secrets in, and never take them out - even with “write any worker” permission
This is easy to implement on durable objects but the trouble is if it’s a userspace worker, anyone with access to the cf account could just replace the secrets worker with one that does have a code path to read secrets