Durable agents

A durable agent is an agent whose state survives the failure of everything around it. The Pi Durable release (2026-10-01, https://earendil.com/posts/pi-durable/) gives the working definition: a harness that “persists ongoing work, conversations, application state through failures, through clients, and through change itself” (that phrasing is IndyDevDan’s at 0:15, and the announcement’s own words are the machinery below). The name matters. Ordinary agents are persistent too, in the sense that a session file remains on disk, and /resume gets you back the transcript. Durability is about the middle of a step, not the end of one.

The machinery, from the announcement:

• Every step the harness runs, a model call or a tool execution, is a task, and a task writes a checkpoint before moving on. A dead process is reopened over the same storage, finds unfinished tasks, and resumes each from its last checkpoint.
• Rerunning is a declared property. A tool marked safe to replay reruns after a crash. A tool without that mark, a deploy, is reported to the model as interrupted, never repeated. See verifiability-and-one-way-doors.
• Submissions carry a request id, so a client that retries after a crash gets the original answer instead of submitting twice.
• Application state lives in typed documents committed in the same transaction as the transcript, so state and history cannot disagree. One shared object, readable by every agent session on the device.
• Compaction runs as a background task while the conversation keeps going. Handoff (reset()) starts a fresh context from a note and keeps the old messages searchable.
• Conversations fork; many run concurrently against one harness; any number of clients attach, watch, and steer.

The consequence 2026-10-12-youtube-indydevdan-pi-durable demos: the agent can be moved off the laptop onto a machine built for agents, closed over, restarted, and still be sitting where the work was left, holding a non-replayable deploy behind an approval gate. Refresh the page and the state is there. Kill the process and it picks up from the checkpoint. That is what makes a proactive agent possible at all, because an agent that starts work without you present must be able to survive not being watched. See inner-outer-loop-agents, harness-engineering.

The durability frame comes from a wider current. anthropic’s Claude products and OpenAI’s ChatGPT agents now run on detached durable substrate exposed to multiple clients,

and the Pi Durable pitch is that the substrate should be yours: “We must own our agents, our agent harness, and now our durable agents” (2026-10-12-youtube-indydevdan-pi-durable, 1:18). See technological-lock-in.

unsourced

IndyDevDan claims Anthropic and OpenAI build products like these “using durable agent primitives like Pi durable” (0:59). The labs’ use of durable substrate is plausible and he does not give a source; the claim as stated (they use Pi durable specifically) is unverified. Recorded as his assertion.

Sources

• Earendil, “Pi Durable”, 2026-10-01: https://earendil.com/posts/pi-durable/ (read in full 2026-10-25)
• Transcript: Inbox/2026-10-12-pi-durable-proactive-agents-DNDYYHsypZg-transcript.txt