Beyond the Harness: Engineering Durable AI Agents on Your Favourite Harness

Note & Mandatory: It's mandatory to both RSVP and fill this Google form https://forms.gle/whpN8ZUJAuoikGcz6 if either is not done, security won't allow at the venue.

Last session we built the SelfEvaluate agent: it reads code, edits it, runs the tests and checks its own work, safely. But its memory lived in InMemorySaver. Restart the kernel, and every paused approval and every step of progress is gone.

This session picks up exactly there. We'll look at what harnesses like Claude Code, Codex, Hermes and Pi actually persist, and why persistent is not durable. Then we'll make our agent survive the real world:

⁠Crash mid-task → resume from the exact step
Tool already ran → never runs twice
⁠Human approval pending → wait days, then continue
⁠Every step → recorded and auditable

We'll compare the open-source durable stack (Temporal, Restate, Inngest, DBOS, Dapr Agents, Mastra) and upgrade the agent live.
New here? No problem. We'll recap the agent in the first 10 minutes.

Laptops open . Hands-on, in person.