The Automation Boundary

What Agents Handle Today

Agents manage the inner loop well. The idea-to-preview pipeline takes a concept from conversation to running container without human intervention. Memory storage and recall happens automatically — agents store learnings and retrieve them when relevant. Skill execution, dev container provisioning, and CI triggering are all agent-driven. The pattern is clear: anything inside a sandbox boundary, agents handle confidently.

What Humans Handle Today

The outer loop is still manual. Production deploys require a human to trigger the workflow and verify the result. Rollback is entirely manual — SSH in, pull the previous image, restart. Convex recovery after a crash needs someone to check the data and redeploy functions. Monitoring triage, security remediation, scaling decisions, and upstream cherry-picks all require human judgement.

Where We Want to Push

The goal is to move production deploys behind a canary validation gate — deploy, run health checks, auto-rollback if they fail. Convex restarts should be self-healing with data integrity checks. Semgrep findings should be triaged automatically, with only genuine issues escalated.

Prerequisites

None of this works until monitoring runs on a server instead of a laptop, trust boundaries are properly defined for agent-initiated deploys, and the rollback path is tested end-to-end. The boundary moves when the infrastructure underneath it is ready.