A thread on why 2026’s hottest AI discipline is engineering its own obsolescence.
2023: “Learn prompt engineering — it’s the job of the future!”
2024: “Prompt engineering is dead. Context engineering is the real skill.”
2025: “Context engineering is table stakes. Flow engineering is what matters.”
2026: “Harness engineering is the new frontier.”
Notice a pattern?
Every 12 months, the AI industry discovers that its models can’t reliably do what they claim to do — and invents a new discipline to compensate.
Prompt engineering was learning to talk to the model. Context engineering was learning to feed the model. Harness engineering is learning to babysit the model.
Each one is more sophisticated. Each one is the same confession: the model can’t figure this out on its own yet.
Let me be clear — harness engineering is real work and it works.
Stripe ships 1,300+ AI-only pull requests per week using harnesses. LangChain jumped from 52.8% to 66.5% on Terminal Bench by changing nothing about the model — only the harness. OpenAI built a million-line codebase where zero lines were written by humans.
These are genuine achievements. But look at what they’re actually proving: the scaffolding matters more than the intelligence.
That’s not the flex people think it is.
The harness engineering thesis goes like this:
Agent = Model + Harness.
The model is the horse. The harness is the reins, saddle, and bit. Your job is to channel this powerful but unpredictable animal in the right direction.
Sounds reasonable. But ask yourself: if you have to build reins, a saddle, a bit, blinders, a training corral, a feedback loop, a sensor array, a planning module, and a context reset protocol just to get the horse to walk in a straight line…
…how intelligent is the horse?
Here’s what harness engineering actually involves in 2026:
- Designing constraints so the agent doesn’t wander off task
- Building feedback loops so the agent can self-correct
- Implementing context resets because the agent loses coherence over time
- Creating handoff artifacts because the agent can’t maintain state
- Writing “golden principles” because the agent keeps making the same mistakes
- Running background cleanup tasks because the agent produces slop
OpenAI’s own Codex team spent 20% of every week cleaning up “AI slop” before they automated the cleanup — with more AI, wrapped in more harnesses.
That’s not intelligence augmentation. That’s error management infrastructure.
The prompt engineering parallel is exact:
In 2023, people discovered that if you wrote “You are a senior developer with 20 years of experience, think step-by-step,” the model performed better. It was almost alchemical. Say the right incantation, get better output.
In 2026, people are discovering that if you decompose tasks across three specialized agents with structured handoff artifacts, context isolation, and iterative evaluation loops, the model performs better.
Same energy. Higher budget. Identical premise: the model doesn’t understand what you want, so you have to engineer around it.
The Bitter Lesson (Rich Sutton, 2019) says that general methods leveraging computation always beat hand-coded human knowledge.
Harness engineers cite this paper constantly. And then they proceed to hand-code human knowledge into elaborate orchestration systems wrapping the model.
The irony writes itself.
Here’s the uncomfortable question nobody in harness engineering wants to sit with:
If a model truly understands causation, plans effectively, maintains coherent goals over time, and learns from its mistakes — what is the harness for?
A model that actually reasons doesn’t need you to decompose its tasks. A model that actually plans doesn’t need a separate planning agent. A model that actually learns doesn’t need golden principles hard-coded into the repo. A model that actually maintains state doesn’t need context resets and handoff artifacts.
Every component of a harness is a confession about a capability the model lacks.
This isn’t a critique. It’s a prediction.
Prompt engineering was a real skill in 2023. It solved real problems. People built real careers around it. And then models got better at understanding intent, and the skill got absorbed into basic literacy.
Harness engineering is a real skill in 2026. It solves real problems. People are building real careers around it. And when models get better at maintaining coherence, planning, and self-correction — the skill will get absorbed into basic infrastructure.
The timeline is the only question.
Think about what actually makes harnesses necessary: models can’t maintain coherent goals over time, can’t reason about why something failed (only pattern-match on the error), and can’t plan multi-step work without drifting.
Those aren’t prompting problems. Those are reasoning problems. And when reasoning gets solved — through causal understanding, durable state, genuine planning — harnesses get thinner. Not unnecessary overnight. But thinner. Then thinner again. Until the harness is just “give it the goal and get out of the way.”
The progression tells you where this is going:
Prompt engineering: Human crafts the perfect input → model produces output.
Context engineering: Human curates the perfect context → model produces output.
Harness engineering: Human builds the perfect system → model operates within it.
What comes next: Human states intent → model builds its own system and operates within it.
Each transition eliminated a layer of human compensation for model limitations. The last transition eliminates them all.
If you’re a harness engineer reading this: you’re not wrong to be doing what you’re doing. The models we have today need harnesses. The value is real. The results are measurable.
But be honest with yourself about what you’re building.
You’re not building the future of AI. You’re building the present of AI — which is a transitional phase where models are smart enough to be useful but not smart enough to be autonomous.
The harness is the bridge. Not the destination.
Every discipline that exists to compensate for a limitation is one breakthrough away from irrelevance.
Prompt engineering lasted about 18 months. Context engineering lasted about 12. Harness engineering is on the clock.
The only question is whether the next paradigm is another layer of human scaffolding — or the moment the scaffolding comes down entirely.
I know which one I’m building toward.
@quackspace.pds.quack.space