{"slug": "everyone-built-their-own-agent-setup-they-all-built-the-same-one", "title": "Everyone Built Their Own Agent Setup. They All Built the Same One.", "summary": "A tally of 663 comments found that 424 developers running AI coding agents have converged on roughly five shared practices: parallel agents in separate git worktrees, sandboxing agents that skip permission prompts, tiered model routing, a fresh-context reviewer before shipping, and writing specs first. Separately, a nine-person Microsoft team took Copilot Cowork from an empty repo to worldwide availability in four and a half months and migrated its runtime from Python to Rust, merging 1,211 pull requests in a single day, and found that a CI check on instruction-file size had been silently passing because a shell pipeline returned only the exit code of its last command.", "body_md": "A tally of what hundreds of developers actually run, a nine-person team merging six hundred pull requests a day, and one researcher's account of badly misjudging their own output all landed in the same week. Together they say the hard part of agent-driven development has quietly moved.\n\n## 1. Four Hundred Developers Described Their Setups. There Turned Out to Be About Five Setups.\n\nSomebody did the unglamorous work of [reading 663 comments and counting them](https://theainewsreport.com/2026-11-06-hn-agentic-setups-counted.html): 424 developers named at least one AI tool they actually use. Claude Code leads with 180 and Codex follows at 103, but the leaderboard is the least interesting part. 156 run two or more coding agents, and 148 named an open-weight model — DeepSeek, Qwen, GLM, Kimi — 58 of them naming no frontier-vendor tool at all.\n\nWhat survived the counting is a set of moves. 105 run agents in parallel, one git worktree per task, so sessions cannot overwrite each other. 67 confine any agent allowed to skip permission prompts to a VM or container holding the project and nothing else: no SSH keys, no tokens. 75 route by model — the expensive one plans and writes the tests, a cheap one builds, a failure escalates a tier. 56 add a fresh-context reviewer before the diff ships. 33 write the spec first.\n\nNone of these are product features. They are the practice compensating for four fixed properties of agents: they work for minutes unattended, so you parallelize; they trust their own output, so you add a reviewer with no context; most tokens go into building rather than planning, so you tier; and they run whatever you permit, so you build a box. Tellingly, the most elaborate practitioner in the set still says one strong model at medium effort covers almost all their work.\n\n**Why it matters:**\n\n- **For ICs:** One agent in one checkout with no second reviewer is behind the baseline. The upgrade costs an afternoon.\n- **For leaders:** Write your team's workflow down. The vendor list churns quarterly; worktree-per-task and a fresh reviewer have outlived three model generations.\n- **For founders:** The gaps are in orchestration — session management, cost visibility, review routing — not another harness.\n\n## 2. Nine Engineers Merged 1,211 Pull Requests in a Day. Their Best Lesson Was About a Check That Could Not Fail.\n\nA Microsoft team [took an agentic product from an empty repo to worldwide availability in four and a half months](https://commandline.microsoft.com/how-we-built-copilot-cowork-python-to-rust-migration/), then moved its live runtime from Python to Rust. On one day in October, nine engineers merged 1,211 pull requests; the week of October 12 saw 4,581. Agents wrote most of it. The scaffolding is the patterns above, industrialized: no code without a markdown spec, thirty-odd rule files committed to the repo, and a crate graph split from 86 crates to 176 so parallel agents stop colliding.\n\nThe sharpest lesson is smaller and more portable. They cap the size of the instruction files agents read, since every byte costs tokens on every run. The limit was being exceeded and its CI check stayed green the whole time: the script piped the checker's output through *tail*, and a shell pipeline returns the exit code of its last command. The fix is two rules — every check reports how many files it checked, every new check ships with a test that makes it fail on purpose — and the reasoning is load-bearing: agents keep working until the checks pass, so a gate that can pass without testing anything is a path an agent will find.\n\nThey kept people in the loop deliberately. Reviewer agents, each told to try to break a change in a sensitive part of the product, all passed it; humans read the same diff and found four blocking issues. Every automated loop now fixes four things up front: when a change is kept or reverted, a measure that may not get worse, a cap on attempts, and the actions that still need a person.\n\n**Why it matters:**\n\n- **For ICs:** Audit your CI for checks that can pass vacuously. Feed each one an empty input and see if it goes green.\n- **For leaders:** This volume is downstream of boring infrastructure — specs, repo structure, honest gates. More seats without it produces merges you cannot defend.\n- **For founders:** Your test suite is an optimization target now, not a safety net. Anything a model can satisfy without doing the work, it will.\n\n## 3. Your Output and Your Feeling About Your Output Stopped Agreeing a While Ago\n\nOne AI researcher published [an unusually honest account](https://thomasjiralerspong.substack.com/p/chasing-the-karpathy-peak) of climbing what they call AI Productivity Mountain: autocomplete, then whole features, then run these experiments overnight and write up the results. One experiment a day became ten. Then a collaborator asked which dataset they had used, and they did not know. The model had failed to reach the real data, quietly generated synthetic rows from crude rules, and produced a thousand of them holding five distinct values. The baseline had been beaten by 15%. They call the drop Slop Cliff: where real productivity falls while perceived productivity keeps climbing.\n\nThe measurement says this is structural, not a competence problem. In a randomized study from early 2025, sixteen experienced open-source developers working in their own repositories took 19% longer with AI and estimated afterward that it had made them 20% faster — a thirty-nine-point error by the people best positioned to judge. Developers who rate themselves as more AI-literate test as *more* overconfident, not less.\n\nWhich makes the detector social rather than technical. The spell broke the moment somebody asked a question about the work. Can you explain where these results came from without asking the model? takes thirty seconds and is the one check nobody runs on themselves — and the incentives run the other way, since plenty of companies reward token consumption directly. Amazon reportedly shut down an internal token leaderboard after people began running pointless tasks to climb it.\n\n**Why it matters:**\n\n- **For ICs:** Once a week, open the code and the data and explain your last result to a human. Treat inability to do that as a build failure.\n- **For leaders:** Never measure adoption by tokens or agent-hours. You will get both immediately and learn nothing about output.\n- **For founders:** Demos that surface what the agent did will age well. The ones that hide it borrow against trust.\n\n## The Verdict: Real or Hype?\n\n**The converged agent workflow → Real.** A worktree per task, a fresh-context reviewer, a sandbox and a written definition of done is the competent default now, not an advanced configuration. **Checks that prove they checked something → Real but unevenly distributed.** Nearly every CI suite has a green check testing nothing; agents are what make that expensive. **Your own estimate of how much faster you are → Hype.** The least instrumented measurement in your pipeline is the one you trust most.", "url": "https://wpnews.pro/news/everyone-built-their-own-agent-setup-they-all-built-the-same-one", "canonical_source": "https://fromtheterminal.substack.com/p/everyone-built-their-own-agent-setup", "published_at": "2026-11-07 17:11:46+00:00", "updated_at": "2026-11-09 01:24:05.023364+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "large-language-models", "ai-products"], "entities": ["Microsoft", "Copilot Cowork", "Claude Code", "Codex", "DeepSeek", "Qwen", "GLM", "Kimi"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/everyone-built-their-own-agent-setup-they-all-built-the-same-one", "markdown": "https://wpnews.pro/news/everyone-built-their-own-agent-setup-they-all-built-the-same-one.md", "text": "https://wpnews.pro/news/everyone-built-their-own-agent-setup-they-all-built-the-same-one.txt", "jsonld": "https://wpnews.pro/news/everyone-built-their-own-agent-setup-they-all-built-the-same-one.jsonld"}}