Latent Space argued on July 14 that AI engineering has quietly folded into mainstream software development, built around agents rather than merely with them. The essay, reported from the AI Engineer World’s Fair by Richard MacManus, traces three years of change since swyx coined the term “AI engineer” in 2023 and lands on five patterns now defining the discipline.
The stakes are practical. Teams that treated 2023-era agent demos as a preview of 2026 production systems built the wrong muscle. What actually shipped this year is infrastructure, not autonomy for its own sake.
Latent Space’s five trends, as reported:
- Harness engineering: the workflows, permissions, context management and evaluation loops surrounding a model now matter as much as the model. Anthropic researcher turned Thinking Machines Lab co-founder Lilian Weng’s 2026 essay on the topic replaces her 2023 focus on agent anatomy with a focus on the system around it.
- Loop engineering: teams split work into an inner loop (the agent executing tasks) and an outer loop (a human or system setting direction and checking results). HumanLayer’s Dex Horthy told the conference “the hype is outrunning the discipline.”
- Software factories and forward deployed engineers: companies including Sierra, Cursor and Warp now embed engineers inside customer organizations to decide which parts of a codebase get automated and where a human reviews the output.
- Coding agents maturing past autocomplete: tools such as Claude Code, Codex, Gemini CLI, Cursor and Warp now explore a repository, edit multiple files, run tests and iterate before returning a result, a capability Vercel is extending into web app frameworks with its new agent tool eve.
- Agent skills as a unit of reuse: markdown files that encode a senior engineer’s workflow, popularized by Anthropic last October, are replacing custom orchestration code, though critics at the event warned of “skills hell” as files go stale between model releases.
The skepticism built into this year’s event is worth noting on its own. Cursor’s Pauline Brunet said enterprise adoption “is still concentrated among early adopters,” meaning the software-factory pitch remains unproven at scale outside a small set of design partners. And the framing fight between “software factories” and “orchestras,” voiced onstage by Conductor’s Charlie Holtz, signals the industry has not settled whether engineers are meant to supervise fleets of agents or stay hands-on with fewer of them.
Compare this to 2023, when the same conference debated whether AutoGPT-style autonomous agents could work at all. The 2026 conversation skipped that question and moved to reliability engineering, which is itself evidence the underlying capability question has been answered, even if the organizational one has not.
For a team’s engineering roadmap over the next ninety days, the actionable shift is to stop scoring agent adoption by autonomy and start scoring it by harness quality: does your setup log inner-loop actions to an outer loop with real evaluation, and do you have an owner for re-auditing skill files every time a frontier lab ships a new model. Skip that ownership question and the skills that worked with last quarter’s model will quietly degrade with the next one.
Latent Space, reported by Richard MacManus and published July 14, 2026.
