Writing — Field Notes

Notes on architecture
& agents.

The reasoning behind the platforms and teams I've led — modernization, coding agents, and building systems people can trust. Posted on LinkedIn and Medium.

Marc Rasmussen Copenhagen 7 field notes 2026
★ FeaturedOpen source · .NET · Context engineering

Coding agents pay for your architecture in tokens

After my Claude Code community talk, many asked for the harness and architecture I'd built. I can't share SimCorp's code — so I distilled the ideas into Brokkr & Eitri, a public repo named for the dwarf brothers who forged Mjölnir while Loki tried to sabotage them. The right patron saints for architecture enforcement.

Same 80 features, built two ways: vertical slices with contract-only seams vs. an identical Clean Architecture twin. Single-feature task: ~1.6k tokens sliced vs ~5.3k layered. More importantly, sliced stays O(1) as the repo grows; layered grows O(n).

The problem with any architecture like this is that it's too easy to break. So everything that can be enforced at compile time is — Roslyn analyzers, a mutation-test canary that asserts the build fails on a violation, and a Claude Code hook that warns agents when they drift across a seam.

3.3×
CHEAPER PER FEATURE
EIT100
TOKEN-BUDGET BUILD GATE
16/16
HEIMDALL · 0 FALSE POSITIVES
Read the full post on LinkedIn ↗

The Archive

6 more notes
Ai
Jul 2026

Orlog vs Mem0: what happens when a memory system is allowed not to know

A memory system for AI agents that would rather admit it doesn’t know something than guess. I ran a 100-run benchmark against Mem0 to see if that trade-off actually holds up. Most AI memory tools share one instinct: never come back empty. Ask a question, they search the store, grab whatever’s…

Read on Medium ↗
Software Development
Jul 2026

Your Architecture Has a Token Bill So I Made the Compiler Collect It

How I turned “our codebase is agent-friendly” from a hope into a build error, with Roslyn analyzers, a token budget, and a little Norse mythology. Here’s an error message you’ve never seen before: The build just failed because a feature got too expensive for an AI agent to understand . Not too…

Read on Medium ↗
AI Strategy
Jun 2026

Your AI roadmap is probably already written

Your engineers just call it tech debt. Ask them one question: “What code do you keep patching because it's really a judgment call?” That's usually where rigid rules keep growing because the underlying decision isn't rule-based. One guardrail: if you need deterministic correctness — money, deadlines, compliance — keep it code.

Read on LinkedIn ↗
Harness
May 2026

Baking harness engineering into agent workflows

Over a weekend I built a harness for my Cowork project — for pull requests, bug fixes, anything that would normally force a context shift. If you're a developer, you know how expensive that is. You can bake harness engineering into Cowork, and it does reduce context switching.

Read on LinkedIn ↗
Vivian · Compliance
Apr 2026

Compliance shouldn't depend on remembering

I interviewed a stack of lawyers about conflict-of-interest checks. They all said the same thing: they email colleagues and wait. So in Vivian, “new case” is a 5-step wizard — conflict checks run automatically in under 3 seconds, and the case doesn't exist in the database until everything is in order. Compliance happens because the system makes it impossible to forget.

Read on LinkedIn ↗
Vivian · Eval
Mar 2026

Vivian hits 80%+ on Supreme Court rulings

Over 8 million tokens later, I have the data. We tested Vivian on 500 Supreme Court rulings across three parameters — the dispositive provision, the reasoning, and the concrete legal conclusion. Ordinary LLMs started around 35–40%, often inventing rulings. Vivian reaches 80%, including EU and tax cases. Each week we re-benchmark against 25 freshly published rulings the models have never seen.

Read on LinkedIn ↗
See all posts on LinkedIn ↗
← Back to terminal Marc Rasmussen · Copenhagen · © 2026