Notes from the team on semantic call graphs, merge conflict architecture, provenance integrity, and what the record of software has to become when most of the code is written by something that isn't you.
AI coding models are not neutral reasoning engines. Their learned assumptions about planning, tools, evidence, review, and completion can become an organization's hidden development lifecycle—and a new form of vendor lock-in.
Read the piece→Tokens and pull requests measure activity. Agent ROI becomes measurable only when intent, effort, verification, review, and cost belong to the same accepted software change.
Read the piece→Ontologies provide deterministic guardrails around probabilistic agents. Durable software change goes further: it preserves the intent, semantic validation, verification, provenance, dependencies, and verifiable state that explain why the code was allowed to change.
Read the piece→A durable task tells the next agent what remains. A durable change record must also preserve why the code exists, what success meant, what evidence was observed, who reviewed the exact state, and what happened when that review failed.
Read the piece→AI coding agents need isolated state, causal dependencies, impact analysis, and provenance—not duplicate repository directories. Causal views provide logical isolation over a shared change graph, while files become a compatibility adapter for compilers and other tools.
Read the piece→Rebase reorganizes Git history by rewriting every downstream commit. Atomic makes grouping a read-time filter over a content-addressed change graph, so you can restructure work even removing a change from the middle of a view, without rewriting a single hash. A patch-theory Git alternative for AI agents, in the Pijul, DARCs lineage.
Read the piece→Git's data model is an ancestry graph: strict, sequential, single-lineage. Great for human edits and reviews. Atomic's is a causal graph plus patch theory, which produces concurrency without coordination and provenance that's a structural property of the graph. Great for agent native source code management where context is critical.
Read the piece→Prompt engineering, then context engineering, then loop engineering. Each shift fixed something real, yet none answered the question that actually matters: how do you know an agent did what it was asked to do? Orchestration wrapped around probabilistic guesses isn't an execution model. What's missing is a proof model.
Read the piece→Before an AI coding agent can complete a task, it has to find the relevant code. In every standard agent today, that means grep, ripgrep, find, or directory listing followed by file reads, burning context before reasoning begins. The problem isn't the search tool. The problem is that the repository has no structural understanding of itself.
Read the piece→Git, the incumbent with 94 percent market share, was designed in 2005 for human developers editing text files. Two decades later, AI agents generate 41 to 46 percent of all new code, and one in four AI-generated pull requests produces merge conflicts that Git's line-level architecture cannot resolve without human intervention.
Read the piece→Long-form engineering and research: semantic indexing, merge conflict architecture, and provenance you can verify. No marketing, no recap threads.