Skip to content

Context Orchestration

Context orchestration defines how context is selected and assembled for each agent task. Explicit rules determine what the task receives each time it runs.

For a PR review, a task can define its context as the issue, research, agreed proposal, implementation diff, notes and engineering standards. The reviewer receives both the change and the evidence explaining what it must achieve, assembled from those specified sources.

Preparing context for a PR review
Context
  • IssueRelevant · included
  • Acceptance criteriaRelevant · included
  • Agreed proposalRelevant · included
  • Implementation diffRelevant · included
  • Engineering standardsRelevant · included
  • Implementation notesRelevant · included
  • Old discussionUnrelated · excluded
  • Abandoned approachUnrelated · excluded
  • Superseded proposalUnrelated · excluded
Outcome
Too littleOptimalToo much

With the same definition and source material, deterministic assembly produces the same context. This makes context selection repeatable; it does not make the agent’s response deterministic.

  • Make task preparation repeatable. Each run uses the same selection rules, so the information supplied does not depend on what an agent happens to remember or carry forward.
  • Control what is included. Specify which sources a task uses and what to select from them, keeping unrelated conversations and material out of its assembled context.
  • Make omissions diagnosable. When a task misses a requirement, inspect what it received and the rules that selected it. That gives you a concrete definition to correct and evaluate.

Atomic workflows provide the task boundary: a defined output and acceptance criteria against which to choose the context. Explicit selection does not guarantee that the definition includes everything the agent needs; evaluations provide evidence for improving it.

Context and access are separate controls. Choosing what to include in a task does not determine everything the agent may read or do through its tools. Factory Governance explains the policies that control that authority.

Runtime evaluations can reveal missing requirements, irrelevant material or findings that were not addressed. That evidence can guide changes to what a task receives. Compare the resulting quality, cost and completion time before adopting a change: fewer tokens alone do not establish a better result. Self-Improvement explains how proposed changes are evaluated.

See Workflow Engine context configuration for the implementation in Nucleus.