SOPs = Slop Cannon

It’s February 2026 and by now your company and leadership is flush from chugging gallons of the AI Kool-Aid. They’re looking for ways to scale that agentic productivity. So they’ve asked you, the AI evangelist to produce the prompts, workflows, tools, and SOPs to unlock the rest of the teams.

The problem? It doesn’t work this way, and you know it.

400cc Technique vs. 1000cc Horsepower

AI is an amplifier, not a savior. If your baseline is brute-forcing non-optimal solutions an agentic workflow with great tooling will lead you to create non-optimal solutions at a higher rate.

You see it every track weekend, a 14 year old kid on a beat up 400cc absolutely smoking grown adults on a brand new 200hp+ superbike. Why? Because the kid understands corner speed, trail braking, looking through corners, and body position. Their technique is incredible, whereas the guy on the liter-bike thinks throwing $40k at a machine buys velocity as he treats every corner like a parking lot and tries to make up for it on the straights.

The 1000cc superbike is your equivalent to the state-of-the-art agentic workflow. It’s impressive on paper and in the right hands, but useless without corresponding pilot’s skills. Loading dozens of MCP tools, hitting auto-model selection, and dumping 20k tokens of context into a prompt, then wondering why the LLM is hallucinating shit patterns and failing a simple refactor is the superbike in Novice group. It’s context pollution masquerading as advanced engineering. The tool is powerful, but your technique is injecting noise into the loop.

Tuning your Harness

Prescribing tools handcuffs your team because it ignores the inner loop. Most agentic tools are just simple orchestration code running an inferencing loop. If you’re waiting for an SOP to tell you how to prompt, you’re treating the model like a magic box instead of a mechanical system you need to tune.

You don’t need an SOP; you need a harness.

A harness is your personal technical setup, the controlled environment where you manage the primitives to maintain backpressure on the agent. In 2026, software engineering isn’t about the lines of code you produce; those are just artifacts of the loop. It’s about the mechanical skill of managing context windows and knowing when the model is starting to crash out.

Standardizing this process across a team is like trying to standardize a rider’s lean angle. It’s physically impossible. Every dev has a different feel for the model’s entropy. By the time a corporate transformation program finishes standardizing your prompts, the underlying model and tooling landscape will have already shifted, making your SOP obsolete.

Final Thought

Stop asking for my prompts or my harness configuration. If you can’t control the basics of the model, you’re not going to make it.

Engineering isn’t following a script; and it never has been, it’s learning new techniques and picking the ones that work for your specific environment. If you’re waiting for an SOP or trying to buy your way into velocity with a prescribed setup, then you should be prepared to compete with agents that can follow and execute the same SOPs.

The SOP is for the agent. The technique is for the engineer.