Just Talk to It – The No-Bs Way of Agentic Engineering (steipete.me)

A solo developer update: the author says agentic engineering now writes almost all their code, and their workflow has been simplified to "just talk to it." They run 3–8 codex CLI agents in a 3x3 terminal grid, each making atomic git commits so history stays clean, and they steer work by thinking about a change's "blast radius"—small, isolated edits vs. large invasive ones. Their everyday model is gpt-5-codex at mid settings (fast, balanced reasoning), often fed brief prompts plus screenshots (used in ~50% of prompts) so the model can find exact code locations. They favor lightweight, parallel windows over branches/worktrees to avoid slow dev servers and rely on escape/status checks to abort or reorient agents mid-run. Why it matters: the post argues codex has materially advanced agentic engineering — larger usable context (~230k tokens vs Claude’s ~156k), more efficient token use, message queuing, and a Rust rewrite that yields much lower latency and memory usage. The author criticizes Claude’s plugins/subagents as patchwork and opaque, preferring direct steering and simple plan prompts; they also note open models (GLM 4.6, Kimi K2.1, Sonnet) are improving but not yet daily drivers. Practical implications: developer DX and tool choice now matter as much as raw model benchmarks — think parallelism, commit isolation, context management, and cost trade-offs (they use multiple subscriptions at ~\$1k/month versus far pricier API usage).