What happens when coding agents stop feeling like dialup? (martinalderson.com)

Coding agents that once felt magical are beginning to act like late‑90s dial‑up: flaky, slow, and prone to retries. Users report reliability issues (notably at Anthropic), and OpenRouter’s small-sample telemetry — caveated as <1% of global traffic and skewed by free “Grok” usage — still indicates an enormous (~50x) jump in token volume. That matters because agentic coding workflows can consume roughly 1,000× more tokens than simple chats, and frontier models today often only generate 30–60 tokens/sec. The result: slow, brittle developer loops when running supervised agents like Claude Code, though experimental stacks (e.g., Cerebras/Gemini CLI) have demonstrated 20–50× higher throughput (~2,000 tok/s), which shifts the bottleneck from the model to the human reviewer. If token throughput rises, agent paradigms will change: fast models could enable unsupervised or semi‑supervised workflows that spawn 5–10 parallel attempts with automated evaluation, delivering richer candidate code in developer‑friendly latency. But infrastructure and economics are major constraints — exploding demand creates an “infinite loop” of consumption, semiconductor scaling is slowing, and providers may introduce new pricing or off‑peak tiers to manage peaks. Practical takeaway: teams should watch throughput and tooling advances closely, design workflows for supervised fallbacks, and be prepared for shifting cost models as agentic development becomes both more powerful and more resource‑intensive.