Selecting Reliability Controls

Classify the Failure Before Changing the Workflow

The controls are not interchangeable. A transient generation miss needs a different response from a missing constraint or a contaminated plan. Retrying an invalid assumption produces repeated guesses; adding context to an overloaded workflow adds more load; continuing past an unresolved high-risk decision turns it into state.

Failure class	What it looks like	Pull this lever
Missing or noisy context	The agent cannot reliably identify the relevant facts, constraints, APIs, or code paths	Retrieval and grounding
Poor work shape	The run has too many dependent decisions, mixes discovery with execution, or keeps re-deciding scope	Shaping the work
Noisy generation	The target is clear, but any single attempt may vary in quality or miss a detail	Sampling and selection
Propagation risk	A wrong intermediate decision would contaminate later work or trigger an expensive side effect	Human checkpoints

These four levers change different variables in the system. Retrieval and grounding raise the quality available to each step. Shaping the work changes step difficulty and dependency shape. Sampling and selection improve a bounded noisy transformation when judgment is reliable. Human checkpoints keep unresolved state from crossing a high-risk boundary.

Reliability levers

Four controls target four failure modes

Diagnose first, then combine controls deliberately. No single lever covers missing facts, poor work shape, generation variance, and propagation risk.

Production workflows rarely pull one lever in isolation. A spec supplies context and a review boundary, validation judges retries and verifies execution, sub-agents change orchestration and isolate context, and context compaction manages what carries over between runs. Match each control to a failure mode — more context, decomposition, retries, or review are not universally safer.