AX over UX
A key insight was that most developer tooling built for humans may be reaching its limits. As agents become the primary producers of code, we are shifting from UX to Agent Experience (AX). Many concepts and abstractions we take for granted such as Git, PRs, branching workflows, repository structures, or even frameworks may simply be legacy constructs optimized for human cognition. But agents do not operate under the same constraints. They can explore multiple solutions in parallel, refactor entire codebases in seconds, and evaluate outcomes programmatically.
CLI vs MCP
This also applies to recent concepts. Several participants observed that agents often perform better when interacting with CLIs rather than MCP integrations. The reasoning is pragmatic: CLIs expose deterministic commands with clear arguments and require less contextual interpretation. A nuance was raised in defense of MCP: the protocol itself may not be the main value. While APIs are typically designed to be stateless, real systems often involve multi-step business workflows across multiple endpoints, which agents can struggle to orchestrate reliably.
The Languages Agents Like
The discussion also touched on how language choice changes when agents write most of the code. Languages with strong feedback loops and static guarantees such as Go and Rust appear particularly well suited for agent workflows. Their compilers and tooling provide clear signals about correctness and optimization, enabling faster agent iteration. TypeScript, however, still has strong advantages in agent ecosystems, mainly because most services and tools agents interact with are built in TypeScript.
The Real Moat May Be Infrastructure
Another major theme was that the real barrier may not be training frontier models, but operating them. Even if a model comparable to the latest generation were open sourced, very few organizations could realistically deploy it. Running frontier systems requires massive engineering around inference optimization, GPU clusters, distributed systems, and memory efficiency, often involving teams spending months optimizing attention, routing, and serving infrastructure. While training data once provided a strong advantage, the real constraints increasingly lie in compute, infrastructure, and operational expertise.