Cincy Deliver 2026

July 24th, 2026
Cincinnati, OH

Compound Velocity: Building Just-In-Time Developer Tooling with AI Orchestration

with Jim Myers

Productivity

93% of software engineers code with AI now. There’s been about a 10% productivity boost with that.

METR famously reported in 2024 that engineers felt 20% faster, but were actually 20% slower.

METR 2026 doesn’t show that slowdown anymore, but the gap in perception remains.

Waterfall per-task can work now.

If we harden both the planning and test phases (with Guides and Sensors, respectively).

Build Contextual Equity, the compound interest of quality in the code and docs.

• Connect
• Encode
• Teach
• Parallelize (watch for burnout from context switching, though)

Principles

• Route to strengths
• Capture the “why”
• Compound every cycle
• Spec before code

For now

Codex is good at Code Review.
Claude has Principal Engineer energy.

Models and their strengths will change like the wind.
Make your harness pluggable and model-agnostic as they emerge.

The speed you feel when coding with AI is local and fragile. It doesn’t last or scale.

• Build just-in-time
• Engineer context
• Compound knowledge

The processes out there were built for how software was made 20 years ago.

Genies are good enough at coding now that when they make mistakes, it’s on you and the context you gave.

Our instinct is to give the model ALL THE THINGS, and let it sort them out. This doesn’t work.

Use a fresh session for every task/step.

Terminology

Agent = Model + Harness

Harness = Guides (docs, md - non-deterministic) + Sensors (tests, pipeline, etc. - deterministic)

Models follow instructions from Guides about 1/3 of the time. We need Sensors to enforce them.

Roles

• Implement
• Challenge
• Review