2 min lesson
The gate between rungs: artifact-verifiability
Imagine this comes up at work: "A customer asks which work they can safely hand to background cloud agents vs keep in the foreground. What single decision rule do you give them?" Start with the practical move.
Step 1 of 2
The gate between rungs: artifact-verifiabilityforeground vs background
What actually lets an org climb a rung is whether the work can be verified through an artifact. That's the foreground-vs-background decision rule and it's the most useful heuristic you can give a customer. Foreground agents are for interactive, ambiguous, security-sensitive or steered work where you make decisions and watch every step. Background / cloud agents are for execution against clear goals and acceptance criteria. The test: if the output is reviewable through an artifact (an image, a recording, logs, a diff), push it async; if it's fuzzier, stay foreground.
Exploring an unfamiliar API; spikes
Security-sensitive or ambiguous changes
UI tweaks you want to watch happen
Refactors with clear acceptance criteria
Migrating a whole branch or a database
Anything whose result is checkable from a diff/recording
Don't spin up a background agent with a one-sentence prompt to 'refactor the whole codebase.' Background work needs heavy upfront structuring (clear goal, acceptance criteria, often a specific ticket pulled in via MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition.). A fuzzy task handed to an unsupervised agent is exactly the reckless move that costs you the room. Verifiability is the gate: no artifact to check against means it isn't ready to go async.
Expect: 'A bank with a strict change-management culture wants to start with autonomous agents in CI. What do you do?' Strong answer: redirect down the ladder. Start with assistive plus governed workflow, prove quality and safety, then earn the pipeline rung. Call out that leading with the top rung in low trust is the classic mistake.