Jonas Templestein
@jonas
Followers
8K
Following
3K
Media
97
Statuses
3K
CEO https://t.co/7dJOmc0va5, prev. cofounder/CTO Monzo, dad of two
Joined October 2009
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.
17
19
198
Are there any remote MCP servers that actually work?! I got so excited by smithery and by the zapier one, but practically nothing works. Is there a platform where I can. 1) do a browser-based oauth with a third party.2) get an MCP server URL and credentials.3) use that MCP server.
2
0
6
Has anyone had any luck using the "official" linear remote MCP server? Any time I try to use it with the openai responses API remote MCP server and include a Bearer token in the Authorization header I get a 401 error. But I know the token is valid because.
1
0
2
Excellent blog post from London's own @viralpickaxe.
At @attio, we made building AI-powered product features as routine as writing to a database. No more fragile prompts - just robust, structured code. Check out my recent post on our 🆕 eng blog (+ a sneaky peek at our homegrown internal tools 👀):.
0
0
2
I started telling my agents they have the same affliction as the protagonist in Memento and it seems to have the desired effect.
Telling the agent to keep track of the work in a markdown is so effective, especially once you automated the loop that it can work for hours.
0
0
6
For example, here's all the code it takes to give the agent a tool to send send itself a reminder event in the future (or on a recurring basis). Giving this tool to an LLM leads to incredible emergent capabilities. All just using human language for "programming". E.g. - "When
In my mind LLM "agents" are just event-driven programs (or as @karpathy would say "Software 3.0" programs). 1) They can receive new events (e.g. a "user message").2) Some events trigger a "computation" (i.e. an LLM request).3) This produces additional events (e.g. an "assistant.
3
2
26
I want to meet London-based AI-maximalist startup founders and set them up with our AI chief of staff. Ideal time to apply is just after you raised a seed round and are looking to "professionalise" (i.e. no bank accounts, accounting, HR setup yet). DMs open - next onboarding.
ℹ️ We’re looking for 2–3 users to trial our operating system for autonomous startups. DMs are open if you're an early-stage, AI-maximalist founder based in London who’s recently raised your first round — or is in the middle of one.
3
8
33
Got deported from the US because I was 5 days out of status between H-1Bs.
2
0
11
“Uses Claude SDK for everything”. This is the way. As much context as/ tools as possible in git. Use filesystem for state management, CLI for local tools and MCP for remote. With coding sandboxes instantly bootable, forkable, cheap and isolated, there is no reason not to run all.
"Jarvis" now has:. - realtime voice that beats current offerings.- natural conversation, context, memories.- contextual awareness, topic shifts, sentiment analysis, associated facts, memories, context.- continual conversational analysis, training, improving and adapting.- tooling.
3
0
6
RT @isamlambert: i’ve never seen a company with low engineering standards turn into one with high engineering standards. if the innate desi….
0
41
0
RT @blankparticle: created this little helper library to add @trpcio styled safe RPC methods with durable objects.this helps us to generate….
0
2
0
To make this concrete: say you want to allow a user to talk to your agent via slack or linear. You could create some shared abstraction (as I did two months ago 🤣), carefully parse all webhooks into schemas, convert them into beautifully formatted user-role input messages etc.
Software 3.0 will be weird. In the end you just serialise stuff into strings / input tokens for an LLM. Any structure your output had before is reduced to a string. This means you probably don’t need that structure to begin with (e.g. elaborate error type hierarchies). You need.
1
1
7
Software 3.0 will be weird. In the end you just serialise stuff into strings / input tokens for an LLM. Any structure your output had before is reduced to a string. This means you probably don’t need that structure to begin with (e.g. elaborate error type hierarchies). You need.
2
1
20
It is so easy to produce reams of mediocre code that I have set a lines-of-code limit for our core logic (excluding tests). Not sure if this is good or bad but it seems to have the desired effect of getting everyone to think a lot more about the code they are asking LLMs to.
1
1
11
Durable object + tanstack query invalidation = 🔥🔥.
Building makeshift convex at home on top of durable objects for @iterate.Its no where near convex level of DX and efficiency, but its a good start
0
0
10
A problem that gets worse with AI coding: Is that anything is possible in programming, but you should do v little of it.
2
0
12