Persistent Systems
Since the emergence of OpenClaw, we are seeing a shift from prompt-based assistants to persistent agents running continuously in the background. Although imperfect, these agents maintain memory, schedule recurring tasks through cron-like loops, and autonomously handle operational workflows. What makes them powerful is not raw intelligence but iteration and accumulated context. Over time they learn preferences, filter noise, and refine workflows without explicit reprogramming. Agents appear most effective when operating in tight loops rather than attempting to solve large problems in one shot, gradually evolving into surprisingly capable autonomous systems.
Static Analysis Meets LLMs
Today many coding agents rely on prompt stacking: when something breaks, the usual approach is to add more context and retry. Hugo shared SkipLabs’ direction: using techniques from static analysis and incremental computation to provide deterministic feedback loops inside the harness of coding agents. Instead of vague signals like “the code failed,” compilers or type systems can return precise diagnostics and dependency tracking. This creates a tighter feedback loop where agents iterate against structured signals rather than repeatedly guessing through prompts.
The Code Explosion
With tools like Claude Code and similar environments, the volume of code produced inside companies is increasing dramatically. Engineers can now generate tens of thousands of lines per day, shifting the bottleneck from writing software to understanding and maintaining it. In extreme cases, some have experienced first-hand teams no longer maintaining a mental model of the system and relying on agents for debugging, and architectural changes. Systems may still function with productivity gains on the surface, but the debugging burden shifts to other teams, making coordination and communication nearly impossible.
Rethinking Interviews
If agents write most of the code, traditional interviews may no longer measure the right skills. Benjamin suggested creating reverse benchmarks where instead of testing the AI we test how well humans collaborate with it. Candidates would work with full agent tooling and be evaluated on how they structure problems, guide the agent, and verify outputs. Observing how someone builds context, decides when to trust the agent, and iterates on results provides a far clearer signal of real-world effectiveness.