Pi Durable
A durable workflow module within the Pi framework that supports state saving and restoration for long-running agent tasks. It incorporates a built-in checkpoint mechanism that enables multi-step AI workflows to resume from the point of interruption rather than restarting from the beginning.
Core Facts
Timeline (last 90 days)
In the Pi ecosystem architecture, Pi Durable replaces the Pi Harness layer; it is likewise built on Pi AI, but provides another interchangeable Harness implementation.
Pi Durable provides an API for managing conversations, but it is not a complete CRUD interface and will not handle all user and conversation management logic on the developer's behalf.
SQLite's transactional properties guarantee the atomicity of Pi Durable's checkpoint writes, so that even a crash during a write will not leave a half-written state.
Pi Durable is worth attention for agent scenarios that require unattended long-running operation and recovery across process crashes; if you are only building an ordinary Coding Agent, the regular Pi Harness SDK may be sufficient.
In Pi Durable, submitting a query returns a Submission ID, and developers can choose to wait for the final result or subscribe to the event stream to read model-generated content and tool calls in real time.
The core of what Pi Durable delivers is a Harness class for developers to import into their own programs; from the application developer's perspective, the way models are called remains unchanged, and only the Harness responsible for long-running operation is replaced.
Pi Durable's checkpoint mechanism periodically writes task execution state (such as whether tool calls have started/completed) to the storage system and associates it with external side effects, thereby avoiding re-executing already-completed external operations upon recovery.
In the demo, the uploader uses Ollama to load the Qwen 3.5 0.8B model, and through the Pi Durable Harness calls the Read tool to read the README file and find the secret phrase.
Pi Durable is currently in an early development stage, and its interfaces are very likely to change in the future.
Initializing the Pi Durable Harness is divided into six steps: configure the model Provider, open the tool registry, set up storage, open the Harness, create or retrieve a conversation, and submit a query.
22 more timeline events
All Facts (20)
Pi Durable is a brand-new runtime that enables agents to run continuously for days, achieves crash recovery through checkpoints, and supports concurrent conversations and multi-interface access.
65%UnverifiedPi Durable provides an API for managing conversations, but it is not a complete CRUD interface and will not handle all user and conversation management logic on the developer's behalf.
50%UnverifiedSQLite's transactional properties guarantee the atomicity of Pi Durable's checkpoint writes, so that even a crash during a write will not leave a half-written state.
50%UnverifiedPi Durable is worth attention for agent scenarios that require unattended long-running operation and recovery across process crashes; if you are only building an ordinary Coding Agent, the regular Pi Harness SDK may be sufficient.
50%UnverifiedIn Pi Durable, submitting a query returns a Submission ID, and developers can choose to wait for the final result or subscribe to the event stream to read model-generated content and tool calls in real time.
50%UnverifiedThe core of what Pi Durable delivers is a Harness class for developers to import into their own programs; from the application developer's perspective, the way models are called remains unchanged, and only the Harness responsible for long-running operation is replaced.
50%UnverifiedPi Durable's checkpoint mechanism periodically writes task execution state (such as whether tool calls have started/completed) to the storage system and associates it with external side effects, thereby avoiding re-executing already-completed external operations upon recovery.
50%UnverifiedInitializing the Pi Durable Harness is divided into six steps: configure the model Provider, open the tool registry, set up storage, open the Harness, create or retrieve a conversation, and submit a query.
50%UnverifiedPi Durable does not automatically read configuration files such as the PI folder or agents.md in the working directory; all configuration must be explicitly loaded when initializing a conversation.
50%UnverifiedA Conversation in Pi Durable roughly corresponds to a Session in Pi Coding Agent, with each conversation having its own system prompt, working directory, tools, and skills.
50%UnverifiedPi Durable is not an out-of-the-box finished application like Pi Coding Agent; it does not provide a ready-made CLI interface, and developers need to implement the interaction interface themselves.
50%UnverifiedIn the Pi ecosystem architecture, Pi Durable replaces the Pi Harness layer; it is likewise built on Pi AI, but provides another interchangeable Harness implementation.
50%UnverifiedPi Durable currently supports three storage backends: SQLite, JSONL file storage, and in-memory storage, and with in-memory storage the data is lost once the process is closed.
50%UnverifiedThe Durable News Agent system uses a layered architecture: at the bottom is a SQLite database, above that is the persistent task capability provided by Pi Durable, above that it is wrapped as an API, and the outermost layer consists of thin clients such as browser, terminal TUI, CLI and Cron.
50%UnverifiedThe core idea of persistent task frameworks (such as Temporal, Restate, Pi Durable) is to write the execution state of each step of a task to persistent storage, so that the task can resume from the breakpoint after a failure and the same task will not be triggered repeatedly.
50%UnverifiedPi Durable's source material does not discuss checkpoint cleanup strategies or compression mechanisms.
50%UnverifiedPi Durable writes to persistent storage on every tool call, which brings additional I/O overhead and storage growth; for Agents with high-frequency calls, JSONL files will rapidly expand, and SQLite will also face write pressure.
50%UnverifiedIn Pi Durable, SQLite suits scenarios that require querying state by condition and guaranteeing transactional atomicity, while JSONL suits log-style appending, post-hoc auditing and quick manual inspection.
50%UnverifiedPi Durable sets the checkpoint granularity to a single task and a single tool call, so that the Agent's execution flow is divided into discrete segments that can be recovered independently.
50%UnverifiedPi Durable's target scenario is always-on Agents reached via Slack, email or Telegram, rather than CLI tools run interactively in a terminal.
50%