Pi 1.0
- Source
- earendil.com
- Author
- sergiotapia
- Date
Pi reached 1.0 with a set of features aimed at keeping agent loops cheap and flexible. Codemode arrives with native MCP support, tools can be loaded on demand instead of all at once, extensions can define virtual models, the harness can warm Anthropic's prompt cache, and system messages can be inserted mid-conversation. Shipped alongside it, Pi Durable is an experimental framework for agents that must survive failures. Latent Space's roundup describes it as a TypeScript port that checkpoints agent state so agents and subagents resume after a crash. It uses swappable storage backends, runs in multiple execution environments under one transcript model, and lets several people steer one agent. Earendil also keeps the codebase small enough for an agent to read, which makes it practical to have your own agent adapt or extend it. Pi Durable is labeled experimental, so treat it as a base to evaluate rather than production infrastructure.
Context
Pi's 1.0 release keeps its extensible coding- framework separate from an experiment in longer-running applications. The companion Pi Durable project addresses tasks and conversations that need to outlive a running process.
Its technical description explains the recovery mechanism: steps save checkpoints, and a replacement process opens the same storage to continue unfinished tasks. An interrupted model request is sent again; an interrupted tool call reruns only when considered safe. Pi Durable was released as an experimental package, so these recovery semantics belong to that companion framework.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Checking sign-in…
Loading comments…