"Pi Durable PROACTIVE Agents: Jev Triggers My ON-CALL Pi Agent" — IndyDevDan
Full transcript: [[2026-10-12-indy-dev-dan-pi-durable-proactive-agents-transcript]].
Why this is in the vault
Durable, detached, multi-client agents are the structural pattern behind the founder's own Mac Mini always-on agent harness and the Jev/TypeSafe classifier integration already in RDCO's toolchain; this video is a live reference implementation of both from the Pi coding agent team, worth benchmarking against.
Episode summary
IndyDevDan walks through "Pi durable" (durable Pi), a new release from the Pi coding agent maintainers (Mario and Armen) that makes agent sessions persist across failures, devices, and clients. He demos a proactive, Jev-triggered incident-response agent that investigates and fixes a production bug without being asked, then tours the durable-agent network's session/state/tasks/tools/extensions UI across his M4/M6 Mac Minis, M5 MacBook Pro, and exe.dev cloud sandboxes.
Key arguments / segments
- [00:02:02]–[00:05:01] Jev classifies a live incoming event stream, detects an incident, and triggers a durable Pi agent that investigates, root-causes, fixes (a classic unsafe null-coalesce bug), and sits "ready for review" until the human approves the deploy — framed as a "proactive agent" that works for you, then waits.
- [00:06:03]–[00:08:01] Durable Pi's core tabs: session (background compaction via a movable compaction mark), state (persistent, shared JSON app state readable/writable by any agent session on a device), tasks, tools, extend.
- [00:09:00]–[00:12:00] Shared application state lets multiple agent sessions on one device read/update the same state object; "tasks" are durable, phase-based repeatable jobs (e.g., workspace snapshot, release-check-with-approval) that can resume from a failed phase instead of restarting.
- [00:14:01]–[00:16:02] Multi-client architecture: agents run detached on a device/workspace and can be attached from a TUI, browser, or any client simultaneously — explicitly framed as how Claude and OpenAI already ship Dots/Claude-through-Slack-style products.
- [00:17:01]–[00:19:00] Tools can be swapped live mid-session (demoed switching an agent to read-only mid-execution); every tool call/model call is its own resumable "task" per the Pi docs (credited to maintainer Earendil).
- [00:20:01]–[00:22:01] Extensions demo: a "Jev damage control" extension puts Jev in front of every bash call to block irreversible commands (blocks
rm -rf, allows it only after the safety extension is explicitly disabled) — paired with a custom system-prompt extension and live-session forking. - [00:24:00]–[00:25:01] Attribution/credit: Dan explicitly frames this release as "taste, judgment, and decision-making," not code volume — thousands of lines of TypeScript, but the product's value is in what the maintainers chose to build.
- [00:26:01]–[00:27:00] Cost data point: Jev classified roughly 6,000 log rows for ~13–16 cents total in the demo session.
- [00:31:00]–[00:32:02] Dan's own adoption plan: all new product work (e.g., a support chatbot) defaults to a durable Pi agent going forward; he is "heavily investing" in proactive agents specifically.
- [00:27:00] Explicit no-sponsorship disclosure: "I'm not sponsored. I don't accept sponsorships on this channel."
Notable claims
- Pi durable is still labeled experimental by its maintainers; Dan says he won't deploy it to production himself until a stable v1 ships [00:25:01].
- Claimed cost: Jev classified ~6,000 incoming log events for 13–16 cents in the session shown [00:26:01]-[00:27:00], and "hundreds" of classifications elsewhere in the demo ran to roughly 1.5–6 cents [00:05:01]-[00:13:00].
- Framing claim: "most agents will be proactive" going forward, and the central agentic-engineering KPI going forward should be "how many agents are doing useful work for you at any point in time," not how many jobs you personally kick off [00:04:01], [00:34:01].
- Three-tier taxonomy introduced: in-the-loop (classic terminal agent), near-the-loop (agent runs a workflow and pings you only for approval/review), out-the-loop (fully proactive, detached, multi-client durable agent) [00:10:01], [00:31:00].
Mapping against Ray Data Co
This is close to a reference architecture for pieces RDCO already runs ad hoc, and its in-the-loop/near-the-loop/out-the-loop taxonomy is the clearest articulation of the observe-vs-act dial discussed in [[2026-10-25-always-on-personal-agent-product-category]]. The Mac Mini always-on agent (LaunchAgent + tmux, iMessage channel) is functionally a single-device, single-client version of what durable Pi generalizes into a device/workspace/agent hierarchy reachable from multiple clients (TUI, browser, curl) — the gap is persistent shared app-state and resumable phased tasks, which RDCO's harness doesn't yet have as a first-class primitive. The Jev-triggered incident-response demo is directly analogous to the "Jev for typed-set problems" pattern already in use (classification/typed-set judgments → Jev over prompt-and-parse) and suggests a concrete extension: a Jev classifier watching a log/event stream to trigger an agent proactively, rather than only being invoked inline during a single agent turn. The "human near the loop" framing (agent completes work, human approves/declines before a non-replayable production action) maps directly onto the founder's existing approval-gated-call discipline (announce before approval-gated actions, send-button stays human-gated) — durable Pi's "safe non-replayable operation" for the deploy step is the same guardrail pattern already codified for autonomous external email send and paper-trade deploys.
Related
- [[2026-09-28-indy-dev-dan-10-levels-of-jev]]
- [[2026-09-21-indy-dev-dan-self-compact-pi-agent-zero-hype-devlog]]
- [[2026-05-18-indy-dev-dan-pi-to-pi-two-way-agent-orchestration]]