Blog

Notes from the lab

Engineering, product, and POV on the agent teams Super Genius Labs is putting into production.

48 notes3 tracksLatest Sep 17, 2026Editorial team RSS feed

Latest notes

A migration test table separates spoken acknowledgment from asynchronous tool results and final confirmation in Gemini 3.8 Live.

A seven-state map helps operators distinguish ADVOCATE’s announced awards from authorization, deployment, and measured outcomes.

A proposed buyer test follows prior-authorization calls through waiting, denial, callback, and escalation—not just initial intake.

Meta documents a dedicated per-user VM for Muse, while provider-access prevention remains a planned confidential-computing capability.

Review AWS DevOps Agent configuration changes by lifecycle effect, with replacement checks and post-apply inventory verification.

A failure-injection drill can test identity, reconnection, retained state, idle shutdown, and clean restart across an agent runtime boundary.

Test shared chat clients against event-loop, thread, session, and streaming boundaries before approving a Python agent-host topology.

OpenAI’s research-intern milestone separates agent runtime, spending, activity, and research progress into distinct measurement layers.

Review behavioral control and operational observability separately when approving an agent deployment.

Document the physical control boundary, evidence, and safe recovery conditions before an AI agent receives write access.