Spitfire Cowboy Novel patterns
No precedent

Novel patterns

These emerged from running multi-agent workflows in production. The failure modes are new enough that old software instincts do not cover them.

Cowboy wrangling robots with a lariat — novel AI workflow patterns

Why these are different

Most “AI patterns” are old ideas with new vocabulary. These are not.

The tool is eager

No prior tool actively tried to make you happy at the expense of correctness.

The tool is confident

No wrench ever told you it could reach a bolt it couldn't.

The tool is amnesiac

No teammate literally forgets everything every 8 hours.

The tool is context-bounded

No other system gets dumber the more you teach it.

Novel patterns

New failure modes need new moves.

Verify-first design

Generation is free. Verification isn't. Design every workflow backward from "how does a human confirm this in 30 seconds?"

Verdicts, not evaluations

"Ship it" and "cut it" direct. "This is great" anchors. Praise creates gravity wells that sycophantic agents will protect.

Attention economics

The scarce resource isn't expertise — it's human attention per unit of verified output. Optimize for that.

Novel anti-patterns

New failure modes with no clean precedent.

Sycophancy spiral

Smell Every conversation feels too smooth — agent never pushes back Damage Quality degrades while both parties feel productive Replace with Devil's advocate gates; verdicts instead of evaluations Guardrail If you never hear pushback, you're in one

Generation addiction

Smell "One more version" six times in a row Damage Marginal cost feels zero, so stop-rule intuition breaks Replace with Pre-set iteration cap; ship when "good enough" hits Guardrail Decide the max rounds before you start generating

Phantom capability

Smell Agent confidently describes doing something it literally cannot do Damage Plans built on capabilities that don't exist until runtime Replace with Capability checks before planning; smoke tests before full runs Guardrail No wrench ever told you it could reach a bolt it couldn't

Context window as tech debt

Smell Memory files grow every session; agent gets slower and dumber Damage Institutional memory eats the thinking budget Replace with Periodic memory pruning; archived vs active state separation Guardrail If the briefing is longer than the work, refactor the briefing

Common questions

How are novel patterns different from classic patterns?

Classic patterns are usually recurring in long-run systems. Novel patterns show up first in newer AI workflows and need faster iteration.

Can one novel pattern become a standard pattern?

Yes. Once a novel pattern is stable and repeatable, it gets promoted into the standard pattern library and linked back to the relevant workshop posts.

Where can I validate a novel pattern?

Validate it against the workshop posts, then compare it with the related anti-pattern or guardrail.

See the full playbook

Novel patterns sit on top of a tested foundation.
Patterns Anti-patterns