Mario and Armin: internal host brief #
Introduction #
In the landscape of AI coding, Pi is arguably the most interesting agent harness to look at today
it gained massive popularity when the AI assistant OpenClaw took the internet by storm. OpenClaw was built on Pi and used it as it’s backbone or spine.
There’s been a lot of developments in Pi since then. A company named Earendill has since acquired Pi and it’s genious creator Mario.
Earendill interestingly was started by Armin, who many will know from the python community. Because he’s made some prolific contributions there like Flask/Jinja/Pygments… just to name a few.
Pi the harness most recently hit 1.0 which is a monumental release for many reasons
But one thing they released on the side with Pi 1.0 is this library that in the last few days has gotten a whole bunch of people excited again. and that’s Pi-Durable.
If you listened to the last Fragmented episode on self-learning harnesses, you’ll notice that evolution and understand why Pi Durable is so so interesting.
But it makes sense to look at Pi-Durable more closely and understand what’s happening there.
I’m extremely fortunate today because both Mario on the show, to help us understand more about Pi Durable….
Mario/Armin, thank you so so much for coming on the show.
In the interest of science/learning
- won’t fangirl/fanboy over the creators of my favorite coding tool today
- will try to extract as much wisdom from you folks.
Topic details and host prompts #
Finding the right primitives ##
While I have you right in front of me….
One of the things that’s impressed me the most about Pi, is how right it gets a lot of the primitives?
Is there a philosophy or some rulebook you follow, that helps you get it right?
- how do you approach getting that api or interface… just right?
I imagine - deciding what belongs in Pi and what belongs in say an extension, is also tought?
The MCP story is super interesting; Famously, Pi didn’t support MCP and then it did…
- Can you walk us through that decision (with that philosohpy)
- Nico’s extension - MCP
- How do you draw that line between what goes into core vs. what stays an extension?
Other examples - where you chose not to add something into Pi Core?
Pi Durable (8–44) ##
And now would love to jump into the main topic. Pi Durable.
Motivation (8–16) ###
Pi already exists, it’s awesome
- What made you wake up one day and decide, you know what…
- actually; we’re going to experiment with this separate harness - Pi Durable?
ok so i’m trying to think about, how you would give the elevator pitch for Pi-Durable:
- what is it?
long running agents
agents in the cloud/multiplayer
- Durability: easier to make long-running in an unattended way | easier to implement recovery, monitoring, etc
- Multi-player
- separating the harness from the compute brings safety and scaling benefits
Durability | concept Temporal + Absurd (16–23) ####
The term “Durable” has a specific meaning in the world of software engineering. What does it mean for an agent?
- everything is a task
- Why checkpoint individual tasks instead of whole turns?
- If replay comes up, ask them to explain it without assuming listeners know workflow internals.
If you’ve worked in a company, that has dealt with some kind of scale, this is actually a common class of problems right? You have a bunch of tasks and you need to throw them into a remote queue for processing.
Famously Temporal is an open source solution.
Why build that engine into Pi-durable itself?
- why not just figure out a way to plug in Temporal
or actually Armin…. you created Absurd, which is a tiny SQL-only library with a very thin SDK to work on top of just Postgres
- that avoids the obv. problem witih Temporal Why not use that?
Did Absurd help inspire some of the designs around Pi Durable?
Durability | Surviving a crash (31–37) ####
One of the other tent-pole features of Pi Durable is
- it’s able to recover really well from crashes
- can keep running for long time.
Can you walk us through that … how is Pi durable able to recover say from a forced process kill, resume and then restart.
What cannot be recovered from task state alone
- workspace files, running commands, credentials?
- plugins
what happens when the outcome of a task is unknown.
- transcript alone cannot establish whether the external action succeeded.
- replay unsafe?
— Recovering the same child.
- If a tool already started a subagent before the crash
- how does the restarted tool find that child rather than spawn another one?
e.g.
A task is suspended for a week, and its plugin changes. What happens when it resumes?
- How is “waiting” encoded
- how does suspension differ from interrupting a running tool? Choose one, not both.
- what is recorded while an agent waits for a human
Multiplayer (23–31) ###
One of the purported features of Pi Durable is a multi-player system. Can you briefly explain this feature of Pi durable.
Multiple clients to the one agent
Pi, cannot connect multiple sessions to it.
- So … My phone, a terminal, and a web UI cannot connect to the same agent.
- cannot stream the same session to all of those?
Does Pi Durable solve this problem?
- How?
- what was the insight that made you believe this could be done
Multiple writers to one session file
- if the parent session from where i spun my subagent owns the process.
- i cannot simply just open a new tab find the session id and take it over
- i mean you can … but it’s then a new separate session
One of my aha moments: I was working on my pi extension; Mario you’ll be happy to know it’s also a subagent extension….
- my requirement was simple: I start a subagent in the background, then want to join it when it gets interesting. Can I watch and give it more context without starting a second agent?
Not very straight forward because of that ownership contract between Pi sessions right?
Ask who owns execution and what survives closing only the client.
Distinguish multiple clients of one agent from multiple writers to one session file.
Self learning/improving harness (37–41) ####
Pi already has /reload. Does Pi Durable add a new dimension to how it’s able
to write things on the fly, and reload it making a self-learning harness.
- Who owns compatibility between recorded task state and new code? Ask whether live reload and restart under a new plugin version have different contracts.
- Ask for a concrete example of safely evolving a running agent. Distinguish hot reload during a live run from recovery under a new code version.
Keep three things separate: loading new code, recovering old work with that code, and evaluating whether the change improves behavior. Ask which are Durable capabilities and which belong to an application built on it. Do not imply that arbitrary code changes are safe or that self-learning is supplied out of the box.
DEEP CUT — Can the agent inspect its own machinery? What does explicit task state let an agent understand or change about its own work that an opaque execution loop would hide? Ask for one demonstrated example and its limits.
Bringing it together: what am I building with Durable? (41–44) ####
Let’s bring it together for us: main features:
- Durable workflows - Surviving Crashes + Long running tasks - resuming tasks
- Enable Multiplayer ecosystems - Both multiple clients to agent + multiple writers to the same session
- Self improving-learning harnesses
Design decision
Why not adapt Pi?
- What was the hardest part of Pi Durable that could not be retrofit
what machinery can I stop building, and
what do I still need to build myself?
Product direction and Earendil (49–53) ###
- Will Pi basically become Pi durable one day?
- What feedback would convince you the task/conversation model is wrong?
Today, it’s a library….
i cannot type pi-durable on my terminal and expect to see a harness
Lead: Will you build your own bot on Durable, or leave that to other developers? Will Pi be replaced with Pi-Durable, if it proves successful?
Closing: What else is Earendil building, and what should listeners watch for?
?? Personal agents (44–49) ###
Quick: why is everyone doing this now, including Mario?
Lead: Why are developers building their own personal agents now, and what does Durable make easier?
5. Rapid fire (53–60) ###
- Will Pi ever ship an official GUI?
- Is Code Mode as Pi’s only tool the future? Do you use Pi that way yourselves?
- What’s the best advice you’ve received as a software engineer?
- What’s the best advice you’d give a software engineer today?
- How do you think software engineering will change?