Armin Ronacher and Ben Vinegar host Geoff Huntley for a 96-minute conversation about coding agents, models that use tools across multiple steps. They discuss preserving state, showing actions, and paying for experiments.

Curiosity has a cost · 4:43

  • Huntley urges developers to learn how agents work. He wants them to explain a coding loop, model inference, and prompt caching on a whiteboard. Prompt caching reuses earlier computation to make repeated requests cheaper.
  • Vinegar pushes back on the moral framing. Developers with families have less time to retrain. Income, identity, and self-worth can all feel threatened at once.
  • The boom-and-bust analogy divides them. Huntley expects scarce expertise to command a premium again. Ronacher expects software demand to grow while the profession loses some of its unusual bargaining power.

Communities can outlast their tools · 17:14

  • The hosts discuss David Heinemeier Hansson’s Rails World keynote and his changing views on AI.
  • Ronacher sees communities preserving their original ideas while changing tools. Rails expressed a way of building software; Ruby was its vehicle.
  • Huntley asks how language design should change when agents write the code.
  • Vinegar compares resistance to Bob Dylan going electric. A community can celebrate an innovator, then resist the next experiment.

Pi Durable took four months · 34:50

  • Fast implementations did not settle the design. Ronacher says the team repeatedly solved individual problems in hours. Finding an interface with the right trade-offs took four months.
  • Pi Durable rethinks the harness: the software that manages a model’s tools, execution, and stored state. It aims to preserve work across restarts and support several conversations at once.
  • A linear chat transcript strains under concurrent work. Examples include summarizing old context while work continues, or five Slack users sharing an agent across branching threads.
  • The framework remains experimental. Earendil’s release article says interrupted tools rerun only when declared safe; otherwise the model handles the interruption.
  • Huntley’s provocative response is to revisit AGENTS.md and skill files after model releases. Old workarounds can become new constraints.

Grow a program while it runs · 42:31

  • Huntley connects an LLM to a live Lisp environment. Lisp supports changing and inspecting a running program. He asks for string reversal, uppercase conversion, and URL fetching with JSON parsing.
  • Definitions accumulate into reusable capabilities. The agent can inspect existing functions, define new ones, execute expressions, and roll back a change. His linked project is Jiti.
  • The product becomes its own development interface. Huntley imagines a staff-only menu where users request changes without opening a coding tool.
  • Lisp is the example, not a required destination. The discussion also points to Smalltalk and a web-app adaptation. Vinegar compares these ideas to musicians sampling forgotten records.
  • Jiti’s rollback covers managed code and data. External effects need separate handling.

Code mode makes tool use harder to read · 59:37

  • Ronacher observes models writing short programs to edit files, even when the harness supplies an edit tool.
  • Code mode lets the model write code that coordinates tools. In Pi’s example, JavaScript runs at the harness layer. A shell command runs in the environment holding the project files; those can be different machines.
  • Compact code can become opaque code. Ronacher compares some generated scripts to “code golf,” the contest to write the shortest solution. Vinegar asks whether that habit leaks into application code.
  • Ronacher suspects training rewards completion more than readability.

Black boxes make verification expensive · 1:05:17

  • Hidden reasoning and hidden actions are separate problems. Ronacher criticizes withheld reasoning traces and tool displays that cannot explain arbitrary scripts.
  • A named file-reading tool can drive a clear progress display. Arbitrary code may need another model to explain it, adding cost and another interpretation layer.
  • Critical work needs checks during execution. Ronacher uses medical research as an example. A confident final message gives too little evidence of what happened.
  • Caching shapes architecture. Huntley asks what developers would build without the cost of invalidating cached prompts. Ronacher sees session lock-in as a barrier to trying different models.
  • Both value open-weight models, whose downloadable parameters allow more control, at a hardware cost.

A $20 adapter meets $15,000 in tokens · 1:18:13

  • Ronacher’s CarPlay experiment exposes the price mismatch. He describes a $20 adapter modified to display a small Pi interface. He values its token use at about $15,000 in API pricing, not a $15,000 cash bill.
  • Vinegar gives similar examples: Blender scenes and 3D-printed replacement parts. A cheap physical print can conceal expensive model work and failed attempts.
  • Huntley describes subsidized subscriptions as a way to build habits that spread from home into work. That is his reading of vendor incentives.
  • Huntley also recalls an experimenter who stopped sleeping while building long-imagined projects, then could not work for months.

Useful technology can have fragile suppliers · 1:29:44

  • Huntley thinks capital commitments pressure labs to absorb functions now sold by specialist providers.
  • Ronacher rejects guaranteed financial success. Some companies can lose even if the technology creates lasting value.
  • Huntley separates his commitment to agent-assisted work from loyalty to one lab. He says he would use open-weight alternatives if current suppliers failed.

Who wants to keep changing software? · 1:31:38

  • Vinegar challenges the audience assumption. Programmers enjoy malleable software, meaning software users can reshape. Many customers simply want a dependable tool.
  • Huntley sees a narrower business demand: make a change without switching tools or waiting for another person. He also says automated outreach makes him seek more human contact.
  • Ronacher argues that public value must reach beyond programmers producing more digital content. He raises scientific discovery as a possible example, not an established breakthrough.