David Ondrej turns Armin Ronacher's Pi workflow into a seven-part tutorial

David Ondrej published a seven-part video walkthrough on September 12th documenting Armin Ronacher's Pi coding-agent workflow, which Ondrej calls Ronacher's "exact agentic engineering workflow." Pi is an extensible, minimal agent harness operated by Earendil Works, whose security documentation states it runs with the launching user's permissions and includes no built-in sandbox. Ronacher, the Flask creator who left Sentry in 2025 to build Earendil, described his Pi approach in a January 31st essay, using TypeScript extensions to collect agent answers, manage local to-do files, review changes in separate session branches and surface touched files, favoring conversational back-and-forth over a fixed plan mode.

David Ondrej turns Armin Ronacher's Pi workflow into a seven-part tutorial Ronacher's Pi setup asks developers to customize a minimal agent harness instead of adopting a vendor's workflow. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: X https://x.com/DavidOndrej1/status/2098872445692149820 Why it matters Pi offers an extensible, minimal agent harness, while its security documentation says it runs with the launching user's permissions and includes no built-in sandbox. David Ondrej @DavidOndrej1 https://x.com/DavidOndrej1?ref=runtimewire published a seven-part video walkthrough on September 12th built around the coding-agent workflow of Armin Ronacher @mitsuhiko https://x.com/mitsuhiko?ref=runtimewire , the Flask creator who works on Pi, the agent harness operated by Earendil Works https://pi.dev/news/2026/5/7/pi-has-a-new-home?ref=runtimewire . David Ondrej's seven-part Pi workflow thread https://x.com/DavidOndrej1/status/2098872445692149820?ref=runtimewire Ondrej describes the clips as Ronacher's "exact agentic engineering workflow." The thread directs viewers to Pi https://pi.dev/?ref=runtimewire , with Pi's open-source repository https://github.com/earendil-works/pi?ref=runtimewire hosted by Earendil Works, before closing with links to PostHog and Ondrej's community. The format packages a working developer's habits as a tutorial and gives the project another distribution channel. Ronacher is an unusually credible protagonist for that argument. He created Flask and contributed to foundational Python tools including Jinja and Pygments before spending roughly a decade at Sentry. He left Sentry in 2025 and began building Earendil https://lucumr.pocoo.org/2025/12/22/a-year-of-vibes/?ref=runtimewire , returning to the open-source developer-tool work that made his name. Pi originated with Mario Zechner https://mariozechner.at/posts/2026-04-08-ive-sold-out/?ref=runtimewire , the creator of the libGDX game-development framework. Zechner brought Pi into Earendil in April 2026 and became an Earendil shareholder. He wrote that Pi's technical direction, roadmap, merge decisions and open-source boundaries would be governed by Zechner, Ronacher and Earendil co-founder Colin. Earendil owns the project, while Zechner remains at its helm. David Ondrej's thread packages Armin Ronacher's Pi workflow as a seven-part tutorial. Ronacher works on Pi, the coding-agent harness operated by Earendil; the supplied sources do not establish the precise sequence by which he began using it or how he joined the project. A workflow that begins with fewer features Pi https://pi.dev/?ref=runtimewire is designed around a small core rather than a prescribed software-development process. Its standard tools cover reading, writing, editing and running shell commands. Developers can add TypeScript extensions, skills, prompt templates and custom interfaces, or ask Pi to create those components for them. Ronacher laid out his approach in a January 31st essay https://lucumr.pocoo.org/2026/1/31/pi/?ref=runtimewire . In that account, he says he uses extensions to collect answers to an agent's questions, manage local to-do files, review changes in a separate session branch and surface files touched during a run. He also says he favors conversational back-and-forth over a fixed plan mode. In the same essay, Ronacher describes Pi sessions as trees. He can branch into a fresh context for a code review or tool repair, bring the findings back and return to the main session. He says his review workflow can begin from a clean context and flag details including newly introduced dependencies before a human reviews the changes. Pi's design also allows developers to produce new tools through the agent itself. Its site tells users to ask Pi to build commands, tools, providers, workflows or interface changes, then reload the modified harness in place. The result can become a personal development environment assembled through repeated work rather than a fixed menu chosen by the vendor. That philosophy appears throughout Pi's product documentation. Pi ships without native MCP support, sub-agents, plan mode, permission pop-ups, built-in to-do lists or background shell execution. Ronacher describes the lack of native MCP support https://lucumr.pocoo.org/2026/1/31/pi/?ref=runtimewire as a product choice tied to Pi's preference for user-built extensions. Minimalism moves the security work Pi's flexibility comes with a clear operational consequence. Pi's security documentation https://pi.dev/docs/latest/security?ref=runtimewire says the agent runs with the permissions of the user who starts it and includes no built-in sandbox. Its tools and TypeScript extensions can read or modify files and run shell commands as ordinary local processes. Earendil recommends putting Pi inside a container, virtual machine, micro-VM or other controlled environment for untrusted repositories and unattended work. Project trust settings govern whether local configuration and extensions load. They do not create a security boundary around model output or shell execution. For experienced developers, that design can be an advantage. Pi avoids presenting a partial permission layer as comprehensive protection, and it lets each team choose the isolation model that fits its infrastructure. It also means copying Ronacher's workflow requires more than installing his preferred extensions. Teams need to reproduce the containment, credential controls and review habits around them. PostHog is selling the managed version of the same ambition Ondrej's thread ends with a link inviting viewers to try PostHog's self-driving product https://go.posthog.com/david?ref=runtimewire . The placement connects Ronacher's personal workflow to a separate commercial pitch: software agents that inspect product signals, investigate problems and prepare pull requests. PostHog says https://go.posthog.com/david?ref=runtimewire its agents combine codebase context with analytics events, errors, session recordings, support conversations and external tools. PostHog is packaging autonomy inside a controlled product workflow, while Pi gives developers primitives for constructing their own version. The two approaches expose the choice forming around coding agents. Platforms such as PostHog can own the data connections, sandbox and review queue. Harnesses such as Pi can leave those decisions to the developer and become more personal with every extension added. The resulting Pi setup is opinionated, disposable and closely matched to one engineer's habits. Pi's design gives developers that degree of ownership, while leaving them responsible for assembling the guardrails.