AI Economics and Context Engineering
Two practical questions sit behind most AI programmes: what will this cost over time, and how do we give agents the context they need without waste? These posts take both seriously.
Start here
- The Box is Open argues that cloud AI pricing is subsidised, and that where AI runs should depend on the workload.
- The Real Cost of Intelligence explains why current billing models make return hard to see.
- The Compass Pattern describes a way to cut the context an agent has to load. The headline saving is modelled from one project, so treat it as an illustration.
- Helping AI Agents Remember covers memory files, handoffs and checkpoints for long projects.
If you want help with this
See AI transformation and support and software development consultancy.
All writing on this theme
-
The Box is Open
: Cloud AI pricing is subsidised and will rise. This post argues for choosing local, on-premises or cloud AI by workload, with open-weight models now viable. -
The Process Storage Gap
: AI agents lose track of multi-step work because they keep process state in chat context. The storage layers they lack, and the workarounds teams use today. -
The Database Blind Spot: Why AI Agents Default to Text When They Shouldn't
: AI coding agents default to markdown and CSV where a database would do. This post shows why, and how to prompt agents towards the right data store. -
AI-Adjacent Tools: What's Worth the Complexity
: Most teams do not need another AI tool. Start with markdown, Git and CLI tools, then use five questions before buying observability, orchestration or MCP. -
Engineering Starts Before the First Line of Code
: Treating Claude Code like a new engineer, with discovery, living docs, task breakdown and governance, delivered an event-driven system without major rework. -
Helping AI Agents Remember: Managing Context and State in Long-Running Projects
: AI agents lose context across sessions. External memory files, handoff notes, checkpoints and recovery steps keep long multi-phase projects on track. -
Beyond Code: How AI Agents Are Reshaping Professional Workflows
: How non-developers can use Claude Code, a command-line AI agent, for meeting notes, reports and research, using slash commands, hooks and a prompt library. -
Task Decomposition for AI Collaboration: The Art of Prompt-Ready Work Boundaries
: Vague AI requests give poor results. How to break user stories into bounded, prompt-ready work units with clear inputs, outputs and standards. -
The Compass Pattern: How Smart Documentation Architecture Saves $0.63M Annually
: Copying docs into CLAUDE.md files makes every AI query costly. The Compass Pattern points agents to existing docs, cutting input tokens by 92% in one project. -
Documentation as Decision History: Creating Better Paper Trails for Teams and AI
: Treat decisions like code changes and write them as pull requests. It stops repeated debates, keeps context and gives AI assistants the history they lack. -
The Three Layers of AI Context: Fixed, Session, and Variable
: Split AI context into fixed, session and variable layers and keep each in a markdown template, so you stop re-explaining yourself and get steadier results. -
The Real Cost of Intelligence: Why Current AI Billing Models Don’t Add Up
: Why per-token, per-plan AI pricing discourages trying more than one model, hides the cost of experimenting and muddies ROI, plus ways to budget by outcome. -
From Agents to Architectures: The Future of AI-Native Systems
: What changes when AI moves from assisting to acting: agent problems with state, tooling and accountability, and what AI-native architecture asks of teams.