AI Hash Pipe

A few weeks ago, I wrote that I thought we were lost. Since then, I have spent 10 days of research, planning and the execution of a new framework to build a map that codifies governance, hardens memory, and contains drift with my LLM agents. I ran a dual sprint using Opus on desktop along with Sonnet in Claude Code. One sprint almost followed the process to the letter. The other drifted so bad, that the retro adversarial review cleanup took longer than the planning and execution of the sprint itself.

Hash pipe is short for hash-enforced pipeline. That ended up being the foundation of the solution that evolved as I worked to develop a QA hardening strategy. The core principle is every step of investigation and evidence is output as syntax (YAML) and not prose. In addition, these steps are hash keyed so at the end of every phase during adversarial or retro review, the keys are checked. If they do not match, the phase is rejected and rerun. So even if the agent got the solution correct, if they don't follow the specific instructions of the process it is rejected.

I am learning that LLMs are really good at matching. They are not thinking. The governance and memory layer is there to make sure they have more tightly controlled patterns to match. That idea has been reinforced in my reading of researchers such as Vishal and Varun Sikka and AI pioneers like Yann LeCun. This process has me rethinking my role as "Operator" in this journey and what does it look like for me to review and sign off on the work the LLMs are doing. My current thinking is to treat my LLM more as a project manager and release train engineer than an architect or product manager. I still need to provide the product roadmap, the context for each epic, all of the templates and memory files, and the most important part, understand the design of what the agent has built in order to know if there is drift in the work over time.

The governance and memory layer is there to make sure they have more tightly controlled patterns to match.

I continue to explore the various tools and services that Anthropic and OpenAI seem to be in a race to release, but at the end of the day my goal is to own my memory and governance layer. I will let those two fight it out over who can provide me with the best orchestration partner. I do not see a future where the agents "make it all work". If anything, I think we are going to experience a period where unchecked pattern matching is going to create untracked architecture, hide bugs, and load technical debt into systems faster than anyone can clean it up.

That is where this ties back to leadership. The job isn't only leading people and the agents they are building. It is coaching people to build orchestrators in the first place and to design the constrained patterns that lead to consistent, successful outcomes. That has expanded my thinking (a side effect of the hash pipe!) and is framing the start of a syllabus instead of a simple, "build the agent that tells you what you don't want to hear". Until then, me and the robots are going to keep passing around the pipe and figure out how best to explore what's around the bend.