10 MINUTES · CANONICAL KNOWLEDGE
Concepts
Build the vocabulary of production AI engineering. Each concept starts with a plain definition, shows what breaks, and links back to the system where it matters.
Checkpoint
A durable snapshot from which an interrupted run can continue.
03 · Durable execution conceptReplay
Reconstruct execution from recorded state and decisions.
03 · Durable execution conceptDurable execution
Progress that survives process and infrastructure failure.
03 · Durable execution conceptIdempotency
Repeating an operation has the same intended effect as doing it once.
04 · Side effects conceptReconciliation
Compare desired and observed state before deciding what to do next.
04 · Side effects conceptLease
Time-limited ownership that must be renewed while an owner is healthy.
05 · Work ownership conceptHeartbeat
A periodic signal that proves a worker is still healthy.
05 · Work ownership conceptFencing token
A monotonic token that prevents stale owners from writing.
05 · Work ownership conceptOrphaned run
Work whose recorded owner is no longer able to make progress.
06 · Failure handling conceptHuman interrupt
A durable pause that hands a decision to a person.
07 · Human control conceptContext compaction
Reduce working context while retaining decision-relevant evidence.
08 · Context conceptArtifact
A durable, inspectable output passed between steps or people.
08 · Context conceptTrajectory evaluation
Judge the path an agent took, not only its final answer.
11 · Evaluation conceptEvaluation judge
A rubric-driven evaluator for behavior that cannot be checked exactly.
11 · Evaluation conceptAgent harness
The control layer around a model: tools, state, policy, budgets, and verification.
12 · Deep agents