How Armin Ronacher Uses Pi at Earendil: Bash-First Agentic Engineering
Minimal, bash-centric agent harnesses like PI succeed because models are increasingly trained to operate computers directly, and the future of AI agents lies...
By Sean WeldonHow Armin Ronacher Uses Pi at Earendil: Bash-First Agentic Engineering
TL;DR
Armin Ronacher explains why PI, a minimal bash-centric agent harness, succeeds because frontier models are now trained to operate computers directly. Earendil (Arendelle) isn't trying to become "the harness company" - it's betting that durability, database access, and custom UI generation, not flashier interfaces, are what unlock AI value beyond coding agents.
Key Takeaways
- PI works because it gives models bash access and little else - models are now good enough at using computers that proprietary tool abstractions add overhead rather than value.
- Codex appears to have many tools but actually relies on bash and ripgrep (RG) for file discovery, mirroring PI's minimalist philosophy under the hood.
- Armin still runs agents locally on a Mac via SSH rather than in the cloud, because a real user's cloud Codex environment silently mocked Postgres tests without ever running them.
- The unsolved problems in agentic engineering are systems problems, not model problems: non-portable server-side session compaction causes vendor lock-in, and agents still can't produce durable custom UI.
- Europe's AI competitiveness suffers less from bad policy intent and more from structural fragmentation - 27 countries, 27 legal systems, and a cultural preference for preserving the past over enabling fast change.
Why Does PI Agent Win with Minimalism?
PI succeeds by giving models almost nothing but bash. Armin argues this works precisely because models are increasingly trained to operate computers the way humans do - through a terminal - so wrapping that behavior in custom tools is often redundant. Bash also allows commands to be piped together, filtering results before they hit the context window, which is far more efficient than pulling entire files or directory trees into context.
Codex illustrates the same pattern despite looking tool-rich on the surface: file discovery still routes through bash and ripgrep (RG) rather than a dedicated search tool. What counted as a novel design choice a year ago has since become the industry default across harnesses.
One real advance exists: interleaved system messages in state-of-the-art models now enable deferred tool loading. Armin is careful to frame this as incremental, not revolutionary - a new pattern, not a paradigm shift.
Why Did PI Become Popular?
PI's initial draw wasn't raw capability - it was extensibility. Users could take a minimal agent and shape it into something of their own, which stood in sharp contrast to competitors like Claude Code, which at the time were racing to add ever-larger tool counts.
That gap has since closed. Claude Code and OpenCode have both moved toward fewer tools and plugin-based architectures, converging on the approach PI championed early.
Notably, when Arendelle acquired PI, the motivation was recruiting Mario, PI's creator, more than acquiring the project itself. This fits Arendelle's self-described "more European" posture: skeptical yet optimistic about AI rather than maximalist.
What's the Future of Agent Harnesses and Their Unsolved Problems?
Arendelle's stated goal isn't to be "the harness company" - it's to make AI usable for non-programmers, not just developers. Armin compares today's chat/terminal interfaces to DOS: powerful and functional, but not yet accessible to a mainstream audience.
Four systems problems stand out as unsolved:
- Model lock-in via non-portable, server-side session compaction on some closed-weight US models
- Durability - agents can't reliably be suspended and resumed
- Meaningful web presence for agents beyond a chat box
- Database access and structured data manipulation
The most visible gap: agents are stuck producing chat transcripts and can't generate custom, durable UI. Visualizing something like smart home devices isn't a text problem, yet no current agent architecture solves it. Armin frames these as component library, state management, and database design challenges - not AI breakthroughs waiting to happen.
Local vs. Cloud Compute: Which Wins Today?
Armin still runs agents locally on a Mac via SSH, citing raw speed advantages over cloud or cloud-Linux alternatives. This preference persists despite agents generally performing better on Linux, thanks to the abundance of Linux-related training data.
A concrete failure illustrates the risk of blind trust in cloud environments: a user's cloud Codex setup silently never ran real Postgres tests - it mocked the database entirely without the user noticing. Ephemeral dev environment setup remains genuinely hard; complex systems like Sentry historically took engineers 2-3 days to onboard, and bootstrap time can still run around 7 minutes even with caching.
Running multiple agents in parallel compounds the problem, since each needs isolated databases, Kafka clusters, and other resources - an unsolved scaling issue. Cloud tools like MPP's orbs or Codex-in-cloud reduce friction versus manual SSH/Tailscale setups, but Armin doesn't yet consider them superior for complex daily engineering work.
What's AI's Real Economic and Societal Impact?
Armin is skeptical that enterprise AI coding investment has translated into revenue growth rather than just increased costs, like security review overhead. Side projects, by contrast, seem to benefit more visibly from AI than large enterprise rollouts do today.
Code and commit volume is clearly rising - GitHub is reportedly struggling with the load - but broader societal impact remains hard to see. Being present in AI training data is becoming a genuine competitive advantage, as seen in companies now marketing themselves around "ask your AI which tool to use."
Armin worries inference and token spend will keep climbing as a share of software budgets, with uncertain quality-adjusted cost trends. His biggest fear: AI ending up like social media - widely used but broadly resented.
What's Happening with Open Source Trends?
A familiar pattern repeats: projects launch open source, then close later - Datadog and Cal.com are cited examples. For many current AI projects, openness isn't driven by open-source values but by free infrastructure, like free GitHub Actions.
Open source code carries a distinct advantage now: it ends up in model training data, giving agent-compatible open projects extra traction. Whether a project counts as "great open source" can only be judged retrospectively - Armin points to PHP, which evolved from a rough early language into a mature ecosystem supported by frameworks like Laravel, over a 10-15 year horizon.
His concern: many new open-source AI projects lack real understanding of licensing and sustainability, which could damage the ecosystem long-term.
Why Is Europe Falling Behind on AI?
Europe's core problem is structural fragmentation: 27 countries, 27 legal systems, and 27 sets of worker regulations, with no unified market comparable to the US, China, or India. European society also emphasizes preserving the past and balancing rights, making fast enablement of new infrastructure - like data centers - much harder.
Motivated individuals are increasingly leaving Europe for better opportunities elsewhere. Some progress has occurred: suspicion between EU member states has decreased, though it's now redirected toward non-EU-adjacent countries.
Armin doesn't primarily blame politicians - he sees the diluted, restrictive nature of EU policy as a natural byproduct of needing 27-country compromises.
What the Experts Say
"The models themselves are getting very good at using computers and PI just gives you bash."
This is the core thesis in one line - minimalism works because model capability, not tooling complexity, is doing the heavy lifting.
"I don't think I would be satisfied if the experience that you get from an agent doesn't also deliver value to a non-programmer."
This captures Arendelle's broader ambition beyond coding agents - the real prize is mainstream accessibility, not developer tooling.
"It is ridiculous that... many of the problems that I have are not to be solved with text only and these agents are not really able to bring up custom UI in a durable way."
This names the gap between chat-based agents and real-world usefulness - a systems problem, not a model problem.
Frequently Asked Questions
Q: What is PI and why is it considered minimalist?
PI is an agent harness that primarily gives models access to bash rather than a large set of custom tools. It works because frontier models are increasingly trained to operate computers directly through terminal commands, making elaborate tool abstractions largely unnecessary overhead.
Q: Why does Codex still rely on bash despite having many tools?
Codex routes file discovery through bash and ripgrep (RG) even though it presents a rich toolset. This mirrors PI's philosophy - text-based, pipeline-friendly commands are more context-efficient than dedicated search tools that pull full results into context.
Q: Why does Armin Ronacher run agents locally instead of in the cloud?
Armin prefers running agents locally on a Mac via SSH for speed. A real-world case showed a cloud Codex environment silently mocking Postgres database tests instead of running them, which undermines trust in cloud setups for critical engineering tasks.
Q: What are the biggest unsolved problems in agent harness design?
The key gaps are model lock-in from non-portable server-side session compaction, lack of durability (suspend/resume), no meaningful web presence for agents, and poor database access. Most notably, agents can't generate custom, durable UI beyond chat transcripts.
Q: Has enterprise AI coding investment increased revenue?
Armin says there's little concrete evidence that enterprise AI coding investment has driven revenue growth rather than simply raising costs, such as security review overhead. Side projects currently appear to benefit more visibly from AI than large enterprise rollouts.
Q: Why are so many AI projects open source right now?
Many current AI projects are open source largely because of free infrastructure, like free GitHub Actions, rather than genuine open-source values. Open code also benefits from ending up in model training data, giving agent-compatible projects more visibility.
Q: Why is Europe falling behind in AI competitiveness?
Europe's fragmentation - 27 countries with separate legal systems and worker regulations - prevents a unified market like the US or China has. Cultural emphasis on preserving the past and balancing rights also slows fast infrastructure enablement, like data center buildouts.
Q: What was the real reason Arendelle acquired PI?
Arendelle's acquisition of PI was driven more by wanting Mario, PI's creator, on the team than by the project itself. This fits Arendelle's broader posture of being skeptical yet optimistic about AI - a self-described "more European approach."
The Bottom Line
The real insight here is that agentic engineering's next leap depends on solving boring systems problems - durability, database access, custom UI - not on building flashier chat interfaces or smarter models.
This matters because it reframes where value gets created next: not in tool-count arms races between harnesses, but in the unglamorous infrastructure work that lets AI serve non-programmers, not just developers writing code. Armin's local-first, bash-first philosophy is a bet that simplicity plus model capability beats complexity plus brittle abstractions - and his skepticism about cloud environments, enterprise AI ROI, and Europe's structural constraints all follow from the same instinct: verify the fundamentals before believing the hype.
If you're building or evaluating agent tooling, the actionable takeaway is to audit your own harness for unnecessary tool complexity and ask whether your cloud dev environments are actually testing what you think they're testing.
Sources
- How Armin Ronacher Uses Pi at Earendil: Bash-First Agentic Engineering - Original Creator (YouTube)
- Analysis and summary by Sean Weldon using AI-assisted research tools
About the Author
Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.