Junk that looks like signal
Spam leads, recycled prompts, AI slop. No specificity filter at the top of the funnel, so noise propagates through every downstream stage.
LARKIN SYSTEMS
Field Note No. 01 — Systems Design in Practice
I design cycles,
not pipelines.
An AI agency case study in nonlinear systems design — and the four reasons most AI workflows quietly fall apart.
- Discipline
- Systems design
- Domain
- Content distribution
- Architecture
- Distributed systems
- Stance
- Self-healing
01 The premise
Most AI workflows are pipelines.
A pipeline assumes perfect inputs, infinite capacity, and zero friction. Real conditions provide none of those.
The dominant pattern in AI automation is the chain. One input, sequential stages, one output. It looks clean on a whiteboard. It works in a demo. It collapses the moment a real signal enters it.
Below is the standard linear pipeline most teams build, and the four failure modes it cannot survive.
Output plateaus under load
API limits, budget caps, human bandwidth. Inputs increase but outputs stay flat. No graceful degradation. Just silent failure.
More options, no decisions
The system keeps generating variations. Nothing ships. Analysis paralysis dressed up as productivity.
One bad action ends it
An auto-sent email burns a relationship. A misclassified lead gets blacklisted. There is no undo for reputation damage, and pipelines have no hard stops.
02 The case
An operator scaling without losing voice.
The naive approach: chain a generation step to a translation step to a posting step. Run it five times in parallel for five platforms. Hope it holds.
It doesn't. Within weeks the voice is gone, the editor is the bottleneck, every platform reads like the same homogenized template, and there is no mechanism to learn from what shipped.
The diagnosis isn't bad prompts. It's bad architecture. Linear thinking applied to a problem that is structurally nonlinear.
The diagnosis is never bad prompts. It is always bad architecture.
03 The build
Side by side—a pipeline, and a loop.
The pipeline
what most agencies would have shipped
- Linear flow — One bad stage stops the whole chain.
- Human in every step — The editor becomes the rate-limiter for the entire system.
- No feedback path — The system never learns from what shipped.
- Voice degrades silently — Cross-platform copies homogenize over time.
- Brittle to inputs — A noisy week of ideas produces a noisy week of output.
The loop
what ships in production
- Cyclical flow — Output of each stage becomes input for the next, then the loop closes.
- Parallel fan-out — One insight expands into platform-native variants—not copies.
- Stage gates with auto-repair — Bad outputs get caught, retried with corrective context, or routed to fallback.
- Backpressure — Engagement data flows back upstream—winners amplified, losers throttled automatically.
- Approval gates — Human judgment at specific checkpoints, not continuously. The system runs; you approve.
The loop is not a fancier pipeline. It is a structurally different object. A pipeline is a chain—every link is a failure point. A loop is a self-healing system—failures get retried, signals get amplified, and the system improves under load instead of degrading.
This is what self-healing means in production. It is not a tagline. It is what happens when you stop borrowing your architecture from factories and start borrowing it from event-driven systems with feedback control.
04 The mechanics
Four principles. None of them new—most teams just skip them.
Direct handoff
Zero-latency handoffs. Output of Stage A is immediately input of Stage B. No manual copy-paste, no context loss, no human as a router. The signal stays clean across the entire loop.
Engagement-weighted retries
Bad outcomes throttle upstream. Low-performing patterns automatically dampen production of similar outputs. The system learns what works from what shipped—without anyone having to write a rule.
Platform-native fan-out
One input, many coordinated outputs. A single insight expands into platform-native variants in parallel. Not copies. Adaptations. Each variant respects the physics of its platform.
Human judgment gates
Humans approve, they don't operate. Judgment lives at specific gates—not in every step. You spend your attention where it changes the outcome, and the system runs the rest of the time.
05 The result
What the architecture sustains.
10xContent outputsame operator, same hours
85%Less manual reviewstage gates handle the rest
0Embarrassing postskill conditions catch them upstream
∞Compoundingthe feedback loop sharpens output the longer it runs
More important than any single number: the system gets better as it runs. Engagement data flows back into the input stage. High-performing patterns get amplified. Low performers get retired. The loop learns without anyone teaching it.
That is the difference between automation and architecture. Automation does the same thing faster. Architecture does the right thing, then learns to do it better.
Figures are representative of production loops built on this architecture. Exact metrics depend on baseline, vertical, and operator constraints.
— The method —
I design cycles,
not pipelines.
Nonlinear systems for AI workflows that have to survive contact with reality.