https://youtu.be/jTlQoTH5ZC0?si=YJG1MN7SyAiBB6HO
This is one of the most interesting shifts in AI engineering happening in 2026.
The short answer is:
- Harness engineering asks: How do we build the best environment around an AI agent?
- Loop engineering asks: How do we build a system that repeatedly directs, evaluates, and improves the agent until the objective is achieved?
Loop engineering does not replace harness engineering so much as build on top of it. Think of them as two layers.
|
Era |
Main question |
Human role |
|
Prompt engineering (20232024) |
What should I ask? |
Prompt writer |
|
Context engineering (20242025) |
What information should the model receive? |
Context curator |
|
Harness engineering (20252026) |
What environment should surround the model? |
System designer |
|
Loop engineering (2026) |
How should the whole process keep running until success? |
Process architect |
Harness engineering
The phrase became popular largely through work around AI coding agents, including discussions by OpenAI, Martin Fowler, and others. The central insight is:
An AI agent is not just a model.
Rather,
Agent = Model + Harness
The harness includes almost everything except the LLM itself:
- memory
- tools
- permissions
- repository structure
- testing framework
- verification
- planning documents
- safety constraints
- logging
- rollback
- monitoring
- context management
If GPT-5.5 is the “brain,” the harness is the entire nervous system and workplace.
Ryan Lopopolo from OpenAI described how engineers increasingly spend their time designing these surrounding systems instead of writing code directly.
An analogy:
- race car = model
- track, pit crew, telemetry, rules, fuel, navigation = harness
A Formula One car without its pit crew is nearly useless.
Loop engineering
Loop engineering emerged only recently, especially after Addy Osmani popularized the term.
The key idea is radical:
Instead of
Human → Prompt → AI
you build
Goal → Loop → AI
The loop automatically performs:
- assign task
- run agent
- evaluate result
- run tests
- gather feedback
- update memory
- create next task
- repeat
until some stopping criterion is satisfied.
So the engineer is no longer prompting.
The engineer is designing the machine that prompts.
The philosophical difference
Harness engineering asks
What surrounds intelligence?
Loop engineering asks
What surrounds thinking itself?
That is a much deeper shift.
A harness is mostly structure.
A loop is process.
Why loop engineering is becoming so important
Today’s frontier models are already capable of long chains of reasoning.
The bottleneck is rarely intelligence itself.
Instead, it is things like:
- forgetting objectives
- stopping too early
- not testing
- not retrying
- losing state
- poor prioritization
- failure recovery
A loop continuously fixes these problems.
Instead of asking GPT:
“Write this feature.”
you ask the system:
“Keep improving this software until all tests pass, documentation is updated, benchmarks improve, and code review succeeds.”
The AI becomes a worker inside an automated workflow.
Does loop engineering surpass harness engineering?
I would say yes, but only in the same sense that operating systems surpassed hardware design.
Harness engineering remains essential.
Without a harness, there is nowhere for the loop to operate.
A useful way to picture the relationship is:
Goal
↓
Loop
↓
Harness
↓
Model
The loop decides what happens next.
The harness decides what the agent is allowed and able to do.
The model performs the reasoning.
Why this matters beyond software
I suspect this idea extends far beyond coding.
You have often described thinking not as linear deduction but as following fragments.
A fragment appears.
You follow it.
It changes you.
Another fragment appears.
You follow again.
That is already a kind of loop.
Not a software loop, but a cognitive one.
Your expression,
“follow the fragments,”
can almost be written as
while (life continues):
notice(fragment)
approach(fragment)
allow(fragment to change you)
discover(next fragment)
Notice that there is no predefined path, only a recursive movement.
That is strikingly close to the emerging philosophy of loop engineering.
The difference is that today’s AI loops optimize toward an explicit objective (“pass all tests”), whereas your “follow the fragments” loop optimizes toward something much harder to formalize: the emergence of meaning, or what you have elsewhere connected with strange attractor and Jacques Lacan’s objet petit a. In that sense, human creativity may ultimately depend on loops whose termination condition is not “task complete” but “a new world has begun to disclose itself.”
