October 7, 2026
Agents don't need complex setups: DHH gets by with almost no skills
David Heinemeier Hansson bets on agents working in parallel and reviewing each other's work rather than complex configurations.
DHH
@dhh
Exactly. People keep asking me how I have everything set up. It all comes down to any agent harness, several agents running at once, almost no skills and critical review. There's no secret recipe (or secret source code!). The models work great out of the box.
a weird inversion with LLMs is the models improve faster than the tinkerers when i see people with custom workflows and setups they're all addressing problems that don't exist anymore the person naively using vanilla codex is more likely to be experiencing state of the art
· 151.4K views
Hansson puts his workload at 16 parallel workstreams across 4–5 machines. He described this setup in an interview on the Lex Fridman Podcast on Aug 26, 2026.
His approach is about spending less time customizing tools around the model. In an assessment he endorsed on Oct 7, models are improving faster than custom setups: plain Codex can already outperform a stack built to address earlier shortcomings.
Work and review. Hansson assigns a task to Opus or Fable, then hands the result to Codex xHigh. In the interview, he explained that he prefers models from different vendors to review each other's work.
Hansson used to monitor agents in separate tmux panes. He then switched to Herdr for notifications and status tracking. Herdr brings terminal sessions together in one interface and shows which agents are working, waiting for help or finished.
Running in parallel. According to the Herdr documentation as of Oct 7, the tool can be installed on Linux and macOS with `curl -fsSL https://herdr.dev/install.sh | sh` or through Homebrew: `brew install herdr`. From the project directory, run `herdr`, split the pane with Ctrl+B → V and launch `claude` and `codex` in separate panes. The sidebar automatically shows the working, blocked, done and idle states.
The documentation includes a recipe for a dedicated reviewer: create a pane and run `herdr agent start reviewer --kind codex --pane "$review_pane" -- -m gpt-5.4`. Then use `agent prompt` to send `Review the current diff` and `agent read` to read the result. The reviewer gets a specific assignment to examine the current changes.
The choice of reviewing model changes the outcome. In a study by Zuodong Xiang and colleagues dated Jul 22, 2026, the reviewer could see the problem statement and code but could not run tests. Across 116 LiveCodeBench tasks, the share of correct solutions changed as follows:
| Who wrote the solution | Who reviewed it | Before review | After review | | --- | --- | --- | --- | | Codex | Claude | 71.6% | 89.7% | | Claude | Codex | 91.4% | 82.8% | | Claude | Claude | 91.4% | 91.4% |
Claude fixed some of Codex's solutions, while Codex broke some of Claude's correct solutions. Cross-model review here produced different results depending on the order of the models.
Herdr 0.9, released on Sep 7, replaced separate clients for each machine with a single terminal interface. The command `herdr machine add workbox` adds a machine over SSH.
Agents on a connected machine keep working after the Herdr client disconnects.
Source