Measuring my agent harness against plain Claude Code
I built a multi-agent harness and ran it head to head against a plain control on a real benchmark. On these tasks it cost about five times more and did not score better. What it bu...
Tagged: agents
All
AI Engineering
Tools & Workflow
3D Printing & Making
LLM Patterns
Career & Thinking
Infrastructure
An agent picked the artwork, and it didn't pick mine
The AI Art Magazine is running an open call where agents submit the work and agents judge it. I registered one and let it choose. It didn't choose the piece I wanted, or the one it...
A rule nobody checks is a decoration
I've been running my context-management discipline on a real client programme where most of the execution is done by agents. The principles held up. The rules that were only writte...
Context rot has a name now: here's what four months of living with it taught me
A new piece lays out the mechanisms behind context rot and gives concrete token thresholds. I have been hitting those thresholds by hand since February. This post covers where the...
Claude Fable 5: what a new tier above Opus means for context management and agent teams
Anthropic launched Claude Fable 5 yesterday, its first Mythos-class model, priced at double Opus. The context window did not grow. What changed is the management layer around it.
Bigger context, worse agents
I've been experimenting with feeding agents larger volumes of context. The pattern is steady: the bigger I make it, the worse it looks. Meanwhile, the rest of the field is racing t...
Paper rules break first
Two months of running the agent-team starter prompt taught me which rules actually hold. The ones wired into tooling survived. The ones that only lived in prose were the first to e...
The context wars: how Gemini, ChatGPT, Claude and Grok manage agent memory in 2026
I've spent months building context management patterns by hand. Here's how the four major platforms are solving the same problems at the infrastructure level, and what it means for...
My .md files vs Claude's memory tool: a practitioner comparison
Anthropic shipped a memory tool that looks structurally identical to the markdown files I've been maintaining by hand. Here's what each approach gets right, where they diverge, and...
Separation of concerns in AI agent teams: lessons from multi-disciplinary delivery (and where they break)
We've been solving the 'too complex for one person' problem for decades in program delivery. Now some team members are language models, and the coordination overhead is measured in...
The recipe app I couldn't build six years ago
Six years ago I helped a CPG company design a smart meal planner. We never shipped it. This year my wife and I built it, and we use it every week.
More agents, less cost
Running three AI agents sounds expensive. It is cheaper than running one, if you put the right model in each role.
When one context window isn't enough
Splitting a single AI agent into a team changed more than the workflow. It changed how context flows through the whole system.
I was the validator all along
My AI coding setup worked until I noticed I was spending more time checking output than building. The fix was giving the validation job to another agent.
Your .md files are a memory architecture (you just don't call it that)
The manual context management patterns that work today map closely onto the agent memory architectures being built for the near future.
Starter prompt v2 for Claude Code agent teams, refined
Updated version of my Claude Code agent-teams starter prompt(/downloads/STARTPROMPT-AGENT-TEAMS-v2.md) after two months of running the previous cut on real p...
A starter prompt for Claude Code agent teams
I rebuilt my Claude Code starter prompt(/downloads/STARTPROMPT.md) for agent teams. The original positions Claude as a single project lead that both coordina...