Durable agents are coming. P Durable is not ready for your stack

The signal. P Durable is not the next coding agent. It is the harness layer the P team extracted for custom apps: it replaces the standard P harness, keeps the P AI model layer, and adds durable state via tasks, tool calls, and checkpoints in SQLite or JSONL instead of a session file that dies with the process. The promise: if the process crashes mid-tool-call, a daemon restarts it from the last checkpoint without a continue prompt.

The demo. The video runs a minimal agent on a 0.8B Qwen model through llama.cpp with read, write, edit, and bash tools. The agent reads a README and returns a hidden secret phrase. A second script, ChatResume, interrupts a task right after a tool call; restarting the script reads the SQLite database and finishes the answer. The demo is small, but it makes the checkpointing claim concrete.

The architecture. P coding agent is an application; P Durable is a building block. You create a conversation with its own system prompt, working directory, and tools, but there is no terminal UI and no .pi directory. AGENTS.md is not auto-loaded; you pass instructions explicitly. The harness sits beside the standard P harness and uses the same P AI package for model access and prompt caching. That makes it an SDK for builders, not a ready-made assistant.

The trade-off. Teams get persistence by building the application around it. You own the gateway, conversation handling, and orchestration. The API is early and likely to change. The video recommends SQLite for production, memory only for experiments, and notes that JSONL storage is not the same as P coding agent's session files. For most engineering organizations this is not an adopt-today tool unless you are already building a custom agent backend.

What's missing. There is no production usage, benchmark, or failure analysis beyond a script that resumes after a manual Ctrl-C. The speaker is clear about the library's behavior, but durability under real crashes, partial side effects, and multi-process concurrency are not covered. The four-emails example is illustrative: four sent, one crash, no verification of what actually happened to the in-flight email. That gap matters for teams deciding whether checkpointing is enough.

The implication. The useful signal is not P Durable itself. Agent tooling is moving toward persistent, restartable runtimes with checkpoints for background agents, the same pattern behind several new always-on agent products. That trend will affect how teams design automation triggered from Slack or email and how they budget for state storage and recovery. The video is enough to track the trend; adoption needs a narrow pilot and patience with an unstable API.