# Pi Durable PROACTIVE Agents: Jev Triggers My ON-CALL Pi Agent

## Executive summary

This video introduces Pi Durable, a significant advancement in agentic engineering that moves beyond reactive agents. Pi Durable enables the creation of durable, proactive agents capable of persisting complex application state, running long-running workflows, and operating autonomously across multiple clients and devices. For build engineers, this means agents can handle entire incident response cycles, perform complex development tasks, and wait for human review (Human Near the Loop) without constant prompting, fundamentally changing how systems are built and maintained.

## Key takeaways

- Proactive Agents (Out-of-Loop): The most valuable agents are proactive; they don't wait for a prompt but are triggered by real events (e.g., a Jev classifier detecting a production incident) and work autonomously, only pausing to wait for human review and approval. This is the future of agentic engineering. (03:10, 04:20)
- Durable State and Persistence: Pi Durable maintains a shared, persistent application state object. This state allows multiple, disconnected agent sessions (across different devices or clients) to read and update the same information, enabling complex, multi-agent debugging and development workflows. (06:30, 07:30)
- Multi-Client and Detached Agents: Agents can be detached from a single UI (e.g., a TUI or local device) and accessed via multiple clients (browser, curl, mobile). This allows agents to operate in the cloud sandbox, making them suitable for enterprise product integration. (09:50, 11:30)
- Durable Tasks and Workflows: Agents can execute repeatable, multi-phased jobs (Tasks). These tasks can incorporate sub-agents that require human approval, facilitating a 'Human Near the Loop' process, where the engineer reviews and approves critical steps like deployment. (07:50, 12:00)

## Technical details

- Pi Durable Architecture: Pi Durable provides persistence for ongoing work, conversations, and application state through failures and client changes. It is designed to be an agent harness that allows engineers to own their intelligence and workflows. (00:45, 01:10)
- Agentic Workflow Demonstration: A Jev classifier detects a production incident in a live log stream, triggering a Pi Durable agent. The agent automatically investigates, performs root cause analysis, writes a fix, and presents it for review, demonstrating a full, automated incident response cycle. (03:10)
- State Management: The system utilizes a persistent state object that can be read and updated by multiple, disconnected agents. This shared application state is critical for debugging and coordinating work across different agent instances. (07:30)
- Agent Tooling and Extensions: The agent's capabilities are enhanced through customizable tools (e.g., `list durable`, `check durable job`) and extensions. Extensions allow for advanced features like Jev Damage Control (blocking irreversible actions) and custom system prompts. (13:00, 14:30)

## Practical implications

- **Incident Response:** Automating the detection, investigation, and fix proposal for production issues, significantly reducing Mean Time To Resolution (MTTR).
- **Product Development:** Building customer-facing agents (e.g., support chatbots) that maintain state and continuity across sessions and clients.
- **CI/CD Pipelines:** Integrating durable agents into build pipelines to manage complex, multi-step deployments that require human review at critical checkpoints.
- **System Ownership:** Shifting control from external, rented intelligence (like cloud APIs) to self-hosted, controllable agent harnesses.

## Topics

Agentic Engineering, Durable Agents, Proactive Agents, State Management, Workflow Automation, Human Near the Loop, Agent Harnessing, Pi Durable PROACTIVE Agent Codebase, Pi Durable Release Notes, Pi Durable Playground Demo

Source: https://www.youtube.com/watch?v=DNDYYHsypZg