Che (@chebuilds) header
Che (@chebuilds) profile picture

dropped out to become a founding engineer @trybriefcase. building agents that close the books

London, England
Joined March 2021
Photo by Che (@chebuilds)Photo by Che (@chebuilds)Photo by Che (@chebuilds)Photo by Che (@chebuilds)Photo by Che (@chebuilds)
Pinned Tweet
I rebuilt a feature 3 times before we finally shipped it. For the past 3 months, I’ve been building agents to review your books and correct issues in one click. The first 2 attempts didn’t work. Attempt 1: one agent per account. We quickly realised that accounts are interconnected: journals touch multiple accounts and these agents weren’t able to interact with each other to share knowledge. Attempt 2: one agent per type of issue This was an improvement, however it still had a fundamental problem: there are hundreds of different problems that could realistically arise. It wasn’t scalable to build agents for each type of problem. Despite there being many possible types of problems, in accounting there are only a finite set of ways to fix those issues, most of it coming down to posting or reversing a journal. The takeaway here was that we should enumarate the fixes rather than the problems. Attempt 3: an orchestrator We ended up with a workflow relatively similar to how a human accountant would review the books. It works like this: 1. Start by setting an expectation for what you believe the balance in each account should be 2. Compare the expectation to the current balance 3. Where expectation and reality deviate, dig in to investigate further One top agent (the orchestrator), spawns subagents (investigators) to look into discrepancies and either try to explain them, or hand off to another agent (the drafter) to propose a fix. The accountant always approves. We shipped it last week. Full write-up below.
2
2
5
174
It seems counterintuitive but automating yourself will benefit your own career. If you manage to automate yourself and transform a role with AI, this is a highly in demand skill. On the other hand, if you sit back and refuse to use AI, you will get automated eventually anyway. So do you want to take credit for it or let someone else?
I told the whole company today that their goal should be to replace themselves with AI. If your job can be automated and you refuse to automate it, you’re actually just making it clear you shouldn't have the job.
31
A year ago my take on building with AI was do as much as possible deterministically in a defined workflow and only fallback to an agent when that fails. I’ve changed my mind. Now I think it’s best to start with an agent with a big scope and codify where the agent continuously fails. The core reason is that workflows hit edge cases that agents don’t. With an AI workflow, you are splitting up steps of judgement and routing based on the output of multiple LLM calls. The workflow can’t go back and fix earlier decisions without extra plumbing. An agent, on the other hand, does this out of the box. Start loose, then tighten where necessary.
1
1
2
53
“sf moves faster than London” Don’t you guys have to wait for the light to turn green to cross the road? Grow up
1
102
Claude code and codex handle long-term memory with md files and grep to retrieve. Cursor removed its memory feature entirely. These memory systems are surprisingly simple. Perhaps too simple. As humans we constantly forget things in order to make room for newer, more relevant things. Forgetting is a feature not a bug. Claude code only deletes memories if the file gets too long or if you ask explicitly. Codex removes unused memories after 30 days. And it’s not just these harnesses. Most production agents use some variation of markdown files and/or search over past chats/documents for memory. But this doesn’t feel like the optimal solution for memory. Imagine having an agent that gets better the more it’s used. That’s what a great memory system unlocks, but I don’t think we’re quite there yet.
3
4
146
Design a url shortener is a classic system design interview question yet not even bitly can do it right. Surely this is a bad design decision to allow expired links to be reused?
1
112
Is this what dreams are made of?
1
37
Hot take you should be banned for posting a hot take and it being the most generic take ever
1
28
Gonna be a while before this issue is fixed if all the engineers are using copilot
46
Is there seriously an almost 50/50 shot Jensen’s gone by end of today? Feels way too high
1
21
Guess I’m just stupid then
10
10 nov: mythos figures out how to run on amd in 3 days 12 nov: Jensen resigns Coincidence?
The Nvidia CUDA moat is gone. Anthropic sees all compute as fungible and it only took Mythos 3 days to figure out how to run itself on AMD hardware.
21
are there any truly great agent memory systems out there? feels like everyone is just using md files
3
1
34
The industries most ripe for AI are the ones where the people in them don’t think AI can help
19
The job market in the UK has 2 very different stories. On one side you have a surplus of graduates all fighting for the same grad schemes, most complaining that they can’t find a job. In the startup world, it feels like there is not enough talent. Not everyone is cut out to work at a startup, but I think the main problem is that young people don’t even realise it’s an option. There’s a lot of opportunity in bridging this gap.
1
64
This has to be the worst iOS update in a while. Why does my keyboard feel stupid and wtf is this ugly thing?
51
If you’re young and not willing to work in person you’re ngmi
29
👋 I’m Che, 23, dropped out of an accounting degree 5 years ago. Now I’m a founding engineer building the UK’s first AI-native GL to automate month-end. Building in London? Drop me a dm, would love to connect!
3
5
203
My first post and legit took me 5 mins to figure out where the “+” button was to make a post haha
What’s classed as vibe coding these days? As a founding eng, I barely ever write code by hand. Does this make me a vibe coder?
40