Earendil has shipped Pi 1.0 alongside an experimental companion package called Pi Durable, a framework designed specifically for long-running, crash-resilient AI agents. Pi 1.0 remains the terminal-based coding agent driven by a single human; Pi Durable is the infrastructure layer underneath it, built for applications where agents must survive process death, run concurrent conversations, and be steered by multiple humans. The architecture is straightforward: a harness opens over a pluggable storage backend (memory, SQLite, or JSONL out of the box), keeps only active transcripts and live tasks in memory, and checkpoints every step. If the process dies — laptop sleep, container redeployment, OOM — a new process opens the same storage and resumes from the last checkpoint. Model requests interrupted mid-stream are retried; tool calls are re-executed if safe, or the model is told they were interrupted. Exactly-once semantics on submissions prevent duplicate work after crashes. Concurrency is handled through conversation forking. One harness runs many conversations simultaneously, each with its own agent configuration — model, thinking level, tools, working directory. A conversation can fork another at any point in its transcript, inheriting the parent's history without copying it. Earendil's own analogy: a Slack channel where threads fork from messages. Each conversation stores its own agent settings, so a reviewer can use a cheaper model with read-only tools alongside the main agent. The extension system makes everything pluggable. An extension bundles system prompt sections, tools, hooks, and tasks. System prompts rebuild before every model request, and changes are recorded positionally in the transcript so forks and restarts see exactly what the model saw. On models supporting mid-conversation prompt changes, only the delta is sent, preserving prompt cache validity. The entire codebase without tests is roughly 15,000 lines — about 150,000 tokens for GPT, 250,000 for Claude. This is deliberate: Pi Durable is built so the agent itself can read and understand the framework it runs on. Storage backends alone account for 3,000 lines, and the storage and execution environment interfaces are kept intentionally small for portability across Bun, Cloudflare Durable Objects, or custom key-value stores. The strategic picture is clear. Earendil is not building another coding agent — it is building the substrate on which agentic applications of any kind can run. The harness pattern (storage + model orchestration + tool execution) is a deliberate infrastructure play. Lessons from Pi Durable flow back into the Pi coding agent, but the framework itself is designed to be general-purpose. The bet is that durability, crash recovery, and multi-human steering become table stakes for production agents, and whoever owns that layer owns the agent runtime. The risk is equally clear: infrastructure frameworks live or die on adoption. Pi Durable is experimental, competing against Temporal, LangGraph, and every other orchestration layer angling for the same niche. The 15,000-line codebase is a feature today and a maintenance burden tomorrow if the community doesn't materialize. But the design choices — pluggable storage, small interfaces, agent-readable source — are sound bets on what production agentic systems will actually need.
