HTTP 402, still reserved
Payments were in the web's original vision, but HTTP 402 is still reserved. AI agents may give us another chance to solve identity and accountability online.
Systems
AI, data, agents, reliability, context, and technical architecture.
Three categories of guardrails
There are only three categories of guardrails to prevent harm from agents. First, relying on hard constraints to only allow certain kinds of behaviour. For example, limiting which tokens can be decoded (structured output) or exposing only a specific set of tools to an agent. Assuming correct implementation, these guarantee certain behaviours won't be possible. Second, the LLM&
Core vs edge agents
As our employees use Claude Code for various tasks, we keep seeing agents repeat the same work. Summarizing a customer's state across Salesforce, email, product analytics, usage data: every agent pulls together the same picture from scratch, every time. Martin Kosk, our enterprise architect, proposed a useful frame for thinking about this: core vs. edge. Edge is the
Claude Code IS the software
In 2024, humans created software. In 2025, Claude Code (and Codex, and Cursor, and all the others) created software. In 2026, Claude Code is the software. It's been clear for a while now that the cost of creating software has come down radically, so software becomes disposable: more like a paper towel than a CNC machine. It hasn&
Invest into context
I have a 14,000-word document I maintain manually, to add into ChatGPT queries I make related to my role as CTO. It covers a lot of ground about Pactum: * An extremely basic description of who we are and what we do * Company strategy, long-term goals, financial model, and progress towards those * Our target market, ICP, personas, value
Vibe-code with stable infrastructure
Vibe-coding produces close-to-unmaintainable code right now. But it is an acceptable trade-off in places where it produces decent results and maintainability matters less -- for example making simple UIs. So maybe a good approach is to vibe-code with stable infrastructure. This riffs off of Facebook's engineering principle "move fast with stable infrastructure"
Negotiating against an LLM
Writing this post I got heavy AI assistance, starting off with my raw notes of the competition. I know it's not up to my usual quality – this was the only way I could find time to share this. Overall I consider the result not good enough, but I will keep experimenting with different kinds of AI assistance if
Good LLM devtools are also good human devtools
What does a well designed developer tool (programming language, library, API, CLI utility) look like -- for an LLM? The more we write code with Copilot and Cursor and chat-assistants, the more important this question is. Off the top of my head, a good tool for LLM programmers: * Chunks, abstracts and simplifies, so making most usage pretty concise -- to reduce
LLM+API vs LMM+UI
The two most famous startups focused on making Agent middlewares seem to be Imbue and Adept. Both companies' goal is to have a large model use a computer effectively, but it is interesting how they seem to bet on two different approaches. * Imbue: use LLMs (text-based) with API-based tools. * Adept: use LMMs (multimodal) with UI-based tools.
LLM latency is linear in output token count
All top LLMs, including all GPT-family and Llama-family models, generate predictions one token at a time. It's inherent to the architecture, and applies to models running behind an API as well as local or self-deployed models. Armed with this knowledge, we can make a very accurate model of what the LLM's response time