Which source wins?
When instructions, project memory, runtime evidence, and user intent disagree, authority needs an explicit order.
AI SYSTEMSRUNTIME DESIGNGOVERNANCE
HARNESS ENGINEERING // SYSTEMS AROUND THE MODEL
The model is not the product.
Context, memory, instructions, tools, routing, state, evaluation, approval, and execution boundaries determine what an AI system can know, do, and recover from.
MODEL
CONTEXTauthority · freshness
MEMORYstate · provenance
TOOLScontracts · permissions
ORCHESTRATIONrouting · handoffs
GOVERNANCEscope · approval
EVALUATIONevidence · behavior
RUNTIMEretries · recovery
The thesis
Prompt quality is not enough.
AI products quickly accumulate prompts, rules, memory, tools, routers, evaluators, retries, approval gates, and automation. Those pieces decide whether the system is merely impressive in a demo or dependable enough for real work.
Harness Engineering is my attempt to treat that surrounding layer as engineering rather than as a bag of prompting tricks.
Current research themes
Where the hard questions live.
The framework keeps circling the same engineering tensions.
What may the model do?
Automation should have bounded authority, deterministic stops, and approval gates where consequences deserve judgment.
Who checks the checker?
Independent evidence matters. The same reasoning path should not be the only authority judging its own output.
What evidence already exists?
Do not install telemetry because the diagram looks incomplete. Measure the real gap before adding machinery.
RULE 001
No new machinery without an observed need.
Use deterministic sources where they already exist. Measure first. Add agents, telemetry, or infrastructure only when an observed failure mode earns the complexity.