HYPERREAL ENTERPRISES LTD
Is AI working for you?
Me running agents like a boss.
I thought I was getting really good at agentic coding.
Then I realised the agents weren't building what I asked for.
More automation brings more chances for agents to go off and build things nobody requested or wanted. When a request is well specified, things go better; when it isn't, the agents make up the details as they go, and you end up redoing the work. Automation won't, by itself, solve the problems you want solved.
Have a look at your logs. This Python file reads your Claude Code and Codex logs on your machine. No network, no language model, standard library only. It is short enough to read before you run it. It will tell you what kinds of request you make, how many credentials are sitting in the logs (it doesn't print them), and it draws a chart like the one below of the work your agents did while you weren't typing.
curl -O https://hyperreal.enterprises/downloads/xiaoxiang-local.py shasum -a 256 xiaoxiang-local.py less xiaoxiang-local.py python3 xiaoxiang-local.py --days 7
Download xiaoxiang-local.py · SHA-256 165b79bfd49559f2025e70fbef141cf8e9fddd160460e8a0dc87c1a8239a9f76
Each bar is a gap in my recorded activity: its width is how long the gap lasted, and its height the tokens agents logged during it.
We’re building agentic harnesses for knowledge work.
Foundation models and coding agents are really fun and empowering. With a bit of convincing, you can get ChatGPT to make you a picture of Goofy walking Pluto in the style of R. Crumb, with a nod to the film ‘They Live’. They can help you come up with ideas, and of course you can use them to write code fast, make websites, and even to prove theorems in Lean. But, no doubt you've noticed, there are somethings they can't do very well.
To do the kind of work we are interested in, we had to build a system in which agents can talk to each other and organise work among themselves, in which the whole history can be searched by free text or by graph queries, in which the flow of work can be watched as it happens, and in which Lean supplies the model for the code that's being written. We were our own first client, and the about page shows some of what came of it. We ready to build systems that suit the requirements of your organisation, starting from the tools you already use.
We can create custom memory systems that you own.
An agent is only as good as the knowledge behind it. We build custom memory systems for AI-assisted work on software and mathematics, where a task depends on many linked facts and decisions. Standard coding agents forget what they are doing, what they've agreed to do, and which questions are still open. On complex tasks they run through their context quickly. They will also gladly build you something sort of like what you asked for, leaving no corner left uncut, unless they are given a clear specification of what to build (but then they need to remember to look at the specification). A memory system gives the agents, and you, something to consult. One example is a transparent record of the work as it happens: what was asked, offered, promised and delivered. We run a research prototype of this on our own conversations (work-in-progress notebook); it is not a product. A system built for you runs on your infrastructure and stays under your control. Contact us to discuss what your organisation needs.
AI can help you realise your vision.
But the AI that everyone's talking about these days may only be part of the story.
There is often a gap between what you think you want and what your work actually does. Agents can widen it, or narrow it. The first step is to find that gap. Then we can help you steer.
HYPERREAL ENTERPRISES LTD is a private company limited by shares, incorporated 25 June 2019 as Company Number 634284 on the Register of Companies for Scotland.