Latent Space published its recap of this year's AI Engineer World's Fair, and it named five trends that defined the year. The whole field “entered a new phase: building systems around agents, rather than just building with agents.”
We've been reading the same people everyone else has — Lilian Weng, Dex Horthy, the Anthropic engineering blog, the Latent Space crew — and building alongside the trend, not ahead of it. What we can add isn't foresight; it's texture. Here is what each of the five looks like from the builder's seat, and the posts where we've been chewing on the same problems.
1. The work moved to the system around the agent
The first trend, in the article's words: the focus shifts from agents to the systems around them. It anchors the point to Lilian Weng's work — her 2023 essay on autonomous agents versus her 2026 essay on harness engineering, which names the “harness” as the system layer around the model that handles planning, tools, memory, and evaluation. It's the same analysis as Karpathy's “LLM as a new operating system” and Anthropic's writing on context engineering.
From where we build, this reads as obvious and hard-won at once: the agent is the easy twenty percent. The system around it — the guardrails, the evals, the maintenance story — is the other eighty.
2. Loops became the control layer
The second trend has a name now: loop engineering. Dex Horthy of HumanLayer — his Research-Plan-Implement method and the “Ralph loop,” which developer Geoff Huntley originally ran as a bash loop overnight — headlined a “Great Loops Debate” at the Fair. Weng's harness is, underneath, the coding-agent loop that stabilized across Claude Code, Codex, and Cursor.
We wrote Loops Are the New Harnesses on exactly this — the loop is the new unit of control. The thing we'd add, from having run them in anger: the loop is also the failure surface. The same construct that gives an agent autonomy is how you wake up to a runaway that billed $1,400 in ninety minutes. Control layer and blast radius are the same layer, and treating them separately is how the incidents happen.
3. Enterprise showed up
The third trend: AI engineering entered the enterprise for real. Industry surveys now put roughly three-quarters of enterprises on a two-year timeline for deploying agentic AI, and the Fair added tracks aimed squarely at CTOs and enterprise teams. The money got loud too — the clearest signal being Chamath Palihapitiya's $135M raise for an “AI-native software factory,” which we unpacked in AI-Native Software Factories.
From where we build, enterprise doesn't mean bigger agents. It means the boring parts stop being optional: ownership, audit trails, the maintenance bill that arrives long after the demo. That's the whole argument of Designing Enterprise Software for AI Maintainers — velocity gets you in the door; durability is what keeps you there.
4. The IDE is turning into the agent
Fourth: coding agents are replacing the IDE as the developer's interface. Cursor 3 rebuilt itself around an agent-first layout; OpenAI's Codex went generally available across the editor, terminal, and cloud; GitHub's coding agent moved into Issues. The primary interface is no longer a file you type into.
The editor didn't die; it moved up a level of abstraction.
5. Skills became the packaging layer
Every agent platform is converging on skills. Anthropic shipped Agent Skills in late 2025 — folders of instructions, scripts, and resources an agent discovers and loads on demand via a SKILL.md file — then opened it as a cross-platform standard. The whole ecosystem now rhymes: AGENTS.md, CLAUDE.md, rules, memories, skills, subagents, hooks. Variations on one idea — giving an agent durable, reusable capability without fine-tuning.
Skills are the packaging layer for judgment. Everything we keep saying about human insight mattering more than raw model horsepower only compounds if that insight is reusable — and skills are how you make a hard-won way of doing something portable across every future agent run.
The same set of ideas
The recap closes by noting that builders are “increasingly working with the same set of ideas: coding agents, harness engineering, designing loops, and orchestration.” That convergence is the actual headline — not any single trend, but that a scattered field snapped to a shared vocabulary in about a year.
Building an agent isn't the interesting part. The system around it — the loop, the harness, the skills, the ownership — is where the craft lives now. Which is the same thing we mean when we talk about building production software: the model does the typing; engineering the system around it is the job.
There is plenty of time left in this year and we are very excited to see what the future holds for agent engineering and how we can use it to build better software for our customers.