10% strategy, 20% culture, 70% execution. A better version shipped today beats a perfect one later.
Jan Riethmayer
Product-minded agentic engineering for AI-native systems that survive production.
I build systems measured by outcomes: faster decisions, less busywork, behavior that actually changes. Agents, evals, and quality gates are the means, not the point.
Urgency is a strength. Iteration is the method.
Interested in every facet, from vision and strategy to product, delivery, and system design. But the hours go where behavior changes: 10% strategy, 20% culture, 70% execution.
70%
Execution
Let data find the current bottleneck, then ship the smallest system that moves it, in days not quarters. Outcomes are the argument, software is the evidence.
20%
Culture
Review loops and quality gates make iteration fast. Curiosity, experimentation, and psychological safety make failure cheap enough to learn from.
10%
Strategy
Enough vision to aim the work, compressed into decisions and constraints instead of decks.
Now
Urgency
The latest iteration shipped better than the last and still not 100%. That is the point: iterate in production, not in planning.
Berlin. Hands on the system, not just the strategy: production judgment over demo theatre.
Agents are useful when the whole system is designed.
The interesting question is not whether an agent can produce code. It is whether the system around it can keep producing useful behavior: with boundaries, tests, evals, observability, recovery, and adoption.
Yellow = the loop. Magenta = fails the gate, straight back into iteration. Shipped work ends in the full stop.
Judgment
What should change, for whom, and why now?
Agent shape
Where does autonomy help, and where should humans stay in the loop?
Quality gate
What tests, evals, and reviews prevent speed from becoming entropy?
Adoption loop
Who uses it, what changed, what failed, and what gets corrected?
01
Deterministic gates
Hooks and pre-commit checks enforce the mechanical rules. Nothing depends on an agent remembering.
02
CI mirrors local
One check contract runs identically on the laptop and in CI. Local green means merge green.
03
Binary verification
Quality gates score on exit codes, not vibes. No LLM ever judges the gate.
04
Tiered quality bars
Verification depth scales with blast radius. A docs edit never triggers the full battery.
05
Multi-bot review
Independent AI reviewers plus a human on every PR, looped until unconditionally green.
06
Context at the work
Every directory carries its own agent instructions. Review bots get path-scoped invariants.
07
Skills as playbooks
Repeatable procedures ship as versioned skills. The skills themselves are eval-tested.
08
Evals with ground truth
Prompts are versioned, runs pin BAML prompt and model, results stay reproducible and comparable.
09
Drift ratchet
A deterministic sweep catches what compiles but rots: dead files, duplication, tangled deps.
10
Boundaries agents keep
Ports and adapters enforced by lint at error level. Testable for humans, navigable for agents.
11
Runtime guardrails
Compliance and PII rules enforced in the tool layer and audit-logged, not just documented.
??
Still learning
This slot stays open on purpose. The next practice earns its place in production.
Built where constraints are real.
Earlybird is the current proof environment: busy users, sensitive data, real decisions, and the need for systems that people can trust. Not a lab, not a demo stage.
Current arena
Venture workflows inside a regulated European fund
Role
VP Engineering at Earlybird
Background
3x founder and CTO; product and engineering led together
Practice
AI-native products, internal systems, technical diligence, quality gates
Field notes from the point where strategy becomes software.
The writing is where the work becomes legible: product choices, architecture patterns, adoption gaps, trust boundaries, failed assumptions, and the corrections that followed.
01Product judgment in technical systems
02Agentic engineering patterns
03AI-native workflows that survive production
04Technical diligence for AI, data, infra, and cyber
05Failure reports, eval loops, and trust boundaries
06What changes behavior, not just what demos well
Open thread
Build with judgment. Ship with discipline.
If you care about product-minded engineering, agentic systems, or the line between useful AI and production theatre, the thread is open. Build notes, failures, and corrections land at writing.riethmayer.com.