The two hosts (Mike and a co-host who works on LLM inference cost optimization) debate whether agents will replace developers, starting from economics. Running 100 Codex agents nonstop for a month reportedly costs ~500k just to build a working multi-agent system. Subscriptions are subsidized roughly 1:10 (a heavy 1.5–2k at API prices). Risks: vendor lock-in, silent model downgrades under load, outages (“Claude is down, go home”), and account bans. Agents need real-time supervision (a vending-machine agent was talked into giving products away), and probabilistic agents add entropy to complex systems, which only humans can order. Examples: a million-line AI-written PR porting Bun from Zig to Rust alarmed the community; a professor found 60% of working LLM-written code irrelevant. Agents excel in code and math because of RL with verifiable rewards. Taste-driven domains lag, so “agents replace everyone” may be a tech bubble. Advice: learn what surrounds the LLM (ops, cost, verification), agent infrastructure (protocols, runtimes, memory), and the hardware layer. Humans become agent orchestrators.
Show: explainable
Concepts: llm-cost-optimization · ai-gateway · ai-verification · ai-guardrails · future-of-software-engineering · llm-reasoning