Notes on building software with AI without going soft: how to delegate the work while keeping the thinking.
the frame
Harness engineering is the spine. Every philosophy is really a bet on one of its components:
- Guides (feed-forward): pseudocode, spec-driven, SPDD, memory and rules
- Sensors (feedback): tests and contracts, evals, formal proof
- Loop (orchestration): human-on-the-loop, multi-agent
- Vibe = no harness (eyeball the output). Formal spec = two at once (its spec is a guide, its verifier is a sensor).
Under all of it, one question keeps you honest: are you keeping the thinking (intent, algorithm, verification) and delegating only the work?
the map
start here
- Harness engineering - the spine. Why we need it, and what each component (guides, sensors, loop) is for. Start here.
- The ladder - the philosophies grouped by the component they plug into, the two scoring framings, and the hybrid. With interactive plots.
- Context & memory architecture - a deep dive on the guides / feed-forward component.
the through-line
The hybrid you land on (pseudocode + tests + light spec) is not separate from the harness. Its tests are the harness’s sensors and its spec is a guide. Choosing how you specify is partly building the environment that checks the work.
the larger question
All of this is linked to a much wider question one I keep circling: what makes a good system?