Person/Peter Steinberger/Blog/25/10/Just Talk to It - no-bs Agentic Engineering

• Just Talk To It - the no-bs Way of Agentic Engineering | Peter Steinberger

  • Header graphic credit https://x.com/thorstenball/status/1976224756669309195
    • 
  • Fantastic overview of the SOTA for Agentic AI Coding
  • He has abandoned Claude/Code for Codex/CLI because
    • Codex CLI has ~230k usable context vs ~156k for Claude Code, and token use is more efficient
    • Codex CLI is written in rust and it's faster
    • the language Claude Code use is bad for his "mental health" - the "Absolutely right's," the "100% production ready," the gaslighting, etc, and he finds that Codex CLI has none of this.
    • #Wow Peter is paying $1k per month for "4 OpenAI subs and 1 Anthropic sub" in order to have "unlimited tokens" and he's got a lot of skin in the game. He says,

      • IMO there's simply not much space between the end user and the model company
        

      • I'm not totally sure what he means, but this was said in a paragraph where he was making the point that if he coded with API calls it would cost "around 10x more," so basically subscriptions are the only way to go.
  • He gives a nice overview of many tools,
    • CursorAI - still has bugs from back in May
    • Auggie - blip on timeline, nobody mentioned again

    • RAG might have been helpful for Anthropic/Model/Claude/Sonnet but gpt-5 is so good at searching that you don't need a separate vector index for your code
      

      • My Note - why is it that people are not talking about vector indexes for documentation? It's as though they think that the code is always enough
    • Open/Weights models like GLM/4.6 and Kimi/K2.1 "are strong contendors that slowly reach Anthropic/Model/Claude/Sonnet/3.7 quality
    • Web tools like Devin, OpenAI/Codex/Web, JulesAI from Google,
    • ConductorAI, Claude Code Terragon, SculptorAI, "and the 1000 other ones. Some are hobby projects, some are drowning in VC money ... None stick."
  • He uses Mac/App/WisprFlow for Speech-to-Text

    • Wispr Flow with semantic correction is still king.
      

  • AI/Coding/Background Tasks
    • He reverted back to Claude/Code for a while because of its support for background tasks, which Codex/CLI doesn't have yet, but now he just uses tmux instead and it can run CLIs in persistent sessions in the background, and there's plenty of knowledge about how to do that in the model, so you just need to say "run via tmux" great #Tip !
  • MCP Servers

    • IMO most are something for the marketing department to make a checkbox and be proud. Almost all MCPs really should be clis. I say that as someone who wrote 5 MCPs myself.
      

      • #Example - steipete/claude-code-mcp: Claude Code as one-shot MCP server to have an agent in your agent. #LOL

    • I can just refer to a CLI/Tool by name. I don’t need any explanation in my Agents.md. The agent will try $randomcrap on the first call, the cli will present the help menu, context now has full info how this works and from now on we good. I don’t have to pay a price for any tools, unlike MCPs which are a constant cost and garbage in my context. Use GitHub/MCP and see 23k tokens gone. Heck, they did make it better because it was almost 50.000 tokens when it first launched. Or use the gh cli which has basically the same feature set, models already know how to use it, and pay zero context tax.
      

    • He introduced me to some Observability tools that were New/To/Me
      • Inngest AI and backend workflows, orchestrated at any scale - a bit hard to tell, but if I squint, it looks like a LangSmith and langgraph/Platform competitor.
      • BetterStack Better Stack - Observability Meets Incident Management Looks like a DataDog / Sentry / Splunk / New Relic competitor
    • He says he uses Chrome/DevTools/MCP to replace Playwright/MCP "as my go-to MCP for web debugging ... I don't need to do it lots but when I do, it's quite useful to close the loop."

      • I designed my website so that I can create api keys that allow my model to query any endpoint via curl, which is faster and more token-efficient in almost all use cases, so even that MCP isn’t something I need daily.
        

        • I've not heard about this AI/Coding/Technique before... I'd love to read a more detailed description
  • Flo Crivello on X: "How I prompt my agents https://t.co/N4bm5e3VVw" / X #LOL #Terrible
    • he mentions that Claude/Code responds well to "yelling" but Codex/CLI does not, and this is better for his mental health
  • #Refactoring
    • New/To/Me
      • jscpd
      • knipDev

    • mentioning modern react patterns and rewriting code, e.g. You Might Not Need an Effect – React #React
      

      • also, uses ESLint's React/Compiler

  • AI/Coding/Technique/Spec-Driven Dev

    • in 2025/06 he was using this The Future of Vibe Coding: Building with AI, Live and Unfiltered | Peter Steinberger
    • Now he says that the technique is not worth the time, and starts in chatgpt and moves to code later.