Inner loop versus outer loop is now the working vocabulary: agents run execution in the inner loop while humans set direction, evals, and feedback in the outer one. OpenClaw's Peter Steinberger and Addy Osmani both describe their job that way.
Dex Horthy's objection to loop hype is precise: Kubernetes is built on control loops too, but those loops are deterministic. An LLM loop has no such guarantee, so the discipline has to come from the structure around it.
Compare Lilian Weng's 2023 anatomy of an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → (planning, memory, tool useA model's ability to call external functions — run code, search the web, edit files — instead of only generating text.Full definition →) with her 2026 agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → essay. The unit of engineering moved from the agent's insides to the workflows, context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →, permissions, evaluation, and persistent state around it.
Anthropic's Thariq Shihipar described models as grown rather than designed, with a capability overhead where the model gets smarter in spiky, unpredictable places. That unpredictability is the argument for building monitoring into agent systems.
Cursor's test for a successful enterprise engagement is blunt: the customer does not switch the agents off after the consultants leave. Adoption is still concentrated among early adopters, so finding internal champions is the real bottleneck.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Why it matters
A concise, well-sourced map of where AI engineering is heading — the shift from building individual agents to engineering the reliable loops, skills, and systems around them — useful for practitioners deciding what to prioritize in their own agentic stacks.
Key quotes
“agents can run much more of the inner execution loop, but that outer loop is still engineering.”
