Summary of "DHH’s new way of writing code"

Context

Interview with David Heinemeier Hansson (DHH), creator of Ruby on Rails and CTO/co‑founder at 37signals. He explains how recent advances in AI “agents” have changed how he and his teams build software and run their businesses (Basecamp, Hey, Fizzy, Amachi Linux distro).

Core technological concepts and tools

Agent-first vs autocomplete DHH contrasts older autocomplete-style AI (e.g., inline suggestions, Copilot) — which can feel noisy — with modern agent harnesses that run tools (shell, browser, web APIs) and act autonomously. He sees the agent-first approach as a decisive productivity unlock.
Agent harnesses / IDEs and models mentioned Examples include OpenCode / Cloud Code (harnesses), Cursor, GitHub Copilot, and frontier models he favors (referred to in the transcript as Opus 45 and Claude). The point: combining frontier models with agent harnesses can produce mergeable-quality code.
Token efficiency and Ruby on Rails Rails is described as particularly token-efficient and well suited to agent workflows right now because agents tend to produce and verify human-readable, idiomatic code.
Unix / CLI philosophy DHH emphasizes building CLIs and small interoperable tools (pipes) so agents can orchestrate cross-product work (e.g., a Sentry → GitHub → Basecamp flow).
Quality, security, and verification As agents scale code creation, static analysis and security tooling (e.g., Sonar/SonarQube) and strict review gates become critical to prevent bad dependencies or outages.

Product features and examples

Hey.com An email product with a “screener” that prevents unsolicited messages hitting your inbox (users approve/deny senders). Used as an example of design-driven product decisions and differentiation from Gmail.
Amachi (Umachi) Linux distro Built on Arch + Hyperland, focused on a beautiful, opinionated default install for developers; attracted 400+ contributors and tens of thousands of installs in six months.
Basecamp The long-lived collaboration/tracking surface for the company. DHH is shipping a Basecamp CLI to open agent access and interoperate with other tools.
Fizzy Referenced as one of 37signals’ recent launches.
P1 optimization project An agent-powered side project that massively improved the floor latency (the fastest 1% of requests) by using agents to explore and implement many small PRs quickly.

How teams and processes are changing

Small-team product discovery New projects often start with a tiny core (commonly 1 developer + 1–2 designers) to explore shape and purpose before scaling execution — now accelerated by agents.
Designers as product leads and implementers Designers at 37signals act as product managers and also implement (CSS/HTML/JS), reducing friction and producing better solutions.
Senior developers gain leverage Senior engineers who can validate agent output, give high-quality directions, and supervise agents see the biggest productivity gains (5×–10×).
Risk to junior roles Companies cannot safely let junior devs ship unchecked agent-generated code; junior roles may need redefining toward mentorship and verification skills.
Faster experimentation & more projects Agents lower the marginal cost of exploring hunches, enabling more exploratory projects, quicker PR triage, and rapid prototyping that previously wouldn’t have been attempted.

Practical workflows and demos mentioned

Agent-first workflow Start a project by instructing an agent, review its draft in the editor, make targeted edits, and commit. DHH uses pane setups with agent sessions and a terminal to iterate.
Autonomous agent demos Agents signing up for accounts, following invitation links, and joining projects — illustrative of agents performing multi-step browser workflows.
Large-scale PR triage An agent processed ~100 PRs in ~90 minutes, classifying which to merge, which required rework, and which to discard — showing agents can dramatically reduce backlog.

Safety, limitations, and risks

Agents still make mistakes DHH will not merge sloppy agent output; there is a high bar for aesthetics, readability, and code style.
Production risk Incidents at large companies show you cannot allow unreviewed agent code to hit critical systems.
Security needs Real-time checks for malicious third‑party packages and other supply-chain risks are essential as agents pull external dependencies.
Uncertain timelines The capability improvements are real but timelines and edge-case behavior remain uncertain (compared to the gradual, unpredictable rollout of self-driving cars).

Career and hiring implications

Peak-programmer thesis The old supply constraint (programmers as the bottleneck) may be changing. More software will be produced, and many cost-center development roles may face downward pressure.
Who will stay valuable Engineers with judgment, product sense, design taste, and the ability to orchestrate/validate agents will be in high demand. Top engineers will be even more valuable because they can extract the most from agents.
Hiring practices 37signals’ hiring remains rigorous (heavy screening, at‑home tests, referrals). DHH advises practicing, shipping, and building a track record rather than expecting shortcuts.

Behavioral and human notes

Craft and aesthetics matter DHH ties beauty and correctness together, arguing that well-crafted interfaces and code produce better outcomes and happier users.
Burnout risk and balance Agents are intoxicating; DHH warns against working harder just because agents enable it — preserve sleep, health, and a sustainable pace.
Enjoyment factor DHH finds working with agents more enjoyable — likening the discovery to the feeling of discovering Ruby — and sees agents enabling more creative, ambitious work.

Guides and recommended experiments

Try it hands-on Use an agent harness on a side project or unfinished hobby project to experience the productivity leap.
Start agent-first on new projects Give an agent a clear task, review its draft, and iterate — use agents as a discovery mode.
Build CLIs and small interoperable tools Follow Unix philosophy so agents can compose functionality across products.
Require senior review for production Enforce human validation and static/security checks (Sonar/SonarQube-style tooling) before shipping agent-authored code.

Sponsors / infrastructure mentions

Static (feature flags / experimentation platform)
WorkOS (enterprise SSO, SAML, directory sync)
Sonar / SonarQube (code quality and security tooling)

Main speakers / sources

David Heinemeier Hansson (DHH) — Ruby on Rails creator, co‑founder/CTO at 37signals
Podcast host/interviewer (unnamed)
Referenced models/companies/tools: Anthropic (Claude / “Opus”), OpenCode / Cloud Code harnesses, GitHub Copilot, Cursor, Sonar (SonarQube), WorkOS, Static, Basecamp/Hey/Fizzy/Amachi

Bottom line: The combination of agent harnesses plus improved frontier models has moved AI from intermittently helpful autocomplete to a practical “agent-first” way of building software. That shift accelerates senior engineers, changes team dynamics, makes CLI/APIs more valuable, increases the number of feasible projects, and raises the importance of craftsmanship, security, and human judgment.