A single day building long-running, self-evolving multi-agent systems on Google's agentic stack. The stated exercise is to launch an autonomous agent team that creates data-driven video campaigns, optimizes live ad bids, and self-patches behind eval gates.

The published curriculum:

• Escape naive loops with graph engineering.
• Long-horizon autonomy and zero-cost pausing.
• Orchestrating A2A and multi-modal "doorbells".
• Agent memory management.
• Self-repair and "prompt medics".
• Streaming real-time event-driven arbitrage.
• Building eval-gated self-evolving harnesses.

Four of the seven advertised topics are threads already open here, which makes this a chance to check local findings against someone else's practice rather than a survey of unfamiliar material.

• Eval-gated self-evolving harnesses is the closest match. The standing question on this side is calibration: a gate that has never run against a known-bad input proves nothing, and a harness that rewrites itself makes that worse, because the thing under test and the thing testing it move together. Worth asking how they calibrate the gate a self-patching agent must pass.
• Agent memory management against agent memory and institutional knowledge and the forgetting work in forgetting and attribution.
• Long-horizon autonomy and zero-cost pausing is a durable-execution claim. The AI Tinkerers Boston notes cover durable execution from the other direction; pausing at zero cost is a strong claim and the mechanism is the interesting part.
• Context-rot survival is the organizers' term for what the memory-decay reading calls decay. Whether it names the same phenomenon is an open question, not an assumption.

1. What is the graph in "graph engineering"? A dataflow graph over agent steps, a knowledge graph the agents read, or a state machine?
2. What does an eval gate assert, and has any of them been run against a deliberately broken agent?
3. Does "zero-cost pausing" mean serialized state at rest, or a runtime that bills nothing while suspended?
4. Is "self-patching" bounded by a contract the agent cannot edit?

To be written. The sections above are the announcement; this one is the note.