Pi Hits 1.0: Codemode Ships, and a Durable Substrate for Long-Running Agents

Two days after reversing its MCP stance, Earendil shipped Pi 1.0 on October 1, and the release completes the arc at unusual speed: Codemode is now native, meaning MCP tools, non-LLM models like TypeSafe’s Jev classifier, and image models all sit behind one orchestrated JavaScript sandbox rather than the LLM’s context window. The release notes describe Pi as a “hardened, minimal, extensible agent harness” already used weekly by hundreds of thousands of people, MIT licensed, installable with a curl one-liner. Alongside it comes the more consequential announcement: Pi Durable, an experimental separate package for long-running agentic applications, built for tasks that outlive a terminal session.

What Actually Shipped in 1.0

Beyond Codemode, the 1.0 list is a checklist of harness maturity: extension support for virtual models, deferred tool loading, cache warming for Anthropic models, and mid-conversation system messages that let prompts and tool configurations change as the transcript evolves. The philosophy statement is restraint, with the team noting features are adopted only after proving themselves, a stance consistent with the MCP reversal published the day before: say no until the integration earns its complexity, then ship it properly. The demo shows what the architecture buys: Pi scripts a summary of a week of commits, routes planning to Claude Opus while GPT handles implementation, and lets Jev decide when to switch between them, with per-model cost breakdowns available in the session.

Pi Durable Is the Open Answer to Dots

Pi Durable is the announcement with the widest implications. Where Pi the terminal agent is session-bound, Pi Durable is a substrate for agentic applications that run long, shipped as npm packages (@earendil-works/pi-durable, pi-ai, chord) and explicitly designed for surfaces beyond the CLI. The team frames it as a realization that some use cases do not fit the terminal shape, sharing Pi’s minimalism and what they call “supermalleability.” Read against this week’s landscape, it is the open-source counterweight to OpenAI’s Dots: always-on agents with dedicated compute and purchase authority on one side, an MIT substrate where you own the execution on the other. The always-on agent market now has a closed flagship and an open standard-bearer, and which one enterprises trust with real authority is the question the next six months answer.

The Restraint Playbook Is the Transferable Lesson

The three-post Pi sequence (reject MCP, reverse, ship 1.0 with Codemode inside three days) is worth studying as a product discipline regardless of the tool. Every capability in 1.0 shipped after a public argument for why it earned its complexity, and the Durable package was spun out rather than bloating the terminal agent that hundreds of thousands of people use weekly. Harness teams at larger companies routinely take quarters to make smaller architectural decisions; Pi’s cadence shows what a small team with public reasoning can do, and the community numbers suggest the transparency is the product feature.

What to Watch

Three things. First, Pi Durable’s path from experimental to stable, and whether anyone demonstrates a Durable agent running continuously against production workloads, which would be the open-source proof that always-on agents can be self-hosted. Second, Codemode’s security posture: the Jev injection study applies to Pi’s sandbox too, and a Codemode sandbox escape would be a high-profile find. Third, whether the virtual-model extension pattern catches on: routing between frontier models mid-task with a classifier deciding when to switch is a cost optimization most teams are still doing manually, and Pi just shipped it as an extension point.