The Intelligence Cycle

A canonical five-stage framework describing how intelligence is produced, from consumer question to analytic product delivered to decision-makers. Foundational to intelligence tradecraft education. Failure can originate at any stage — and failures at one stage compound downstream.

The cycle is the closest structural parallel in the intelligence-analysis literature to an LLM agentic pipeline, which makes it a useful framing device for the wiki’s central thesis. Each stage has direct LLM analogs and corresponding LLM-specific failure modes.


The Five Stages

#StageWhat happensCommon failure modes
1DirectionConsumer sets the intelligence question / requirementNarrow framing; ambiguous question; mismatched scope
2CollectionInformation is gathered from sources (HUMINT, SIGINT, IMINT, OSINT)Single-source dependence; collection-strategy gaps; source-bias
3ProcessingRaw collection is validated, translated, structuredFiltering errors; mis-translation; misclassification
4AnalysisAnalysts extract insight, evaluate hypotheses, form judgmentsCognitive bias (see Cognitive Bias); mind-set lock-in; premature closure
5DisseminationFinished intelligence is communicated to decision-makersPolicymaker rejection of accurate intel; format inappropriate to consumer; over-confident language

LLM Agentic Pipeline Parallel

The five-stage cycle maps remarkably cleanly onto modern agentic LLM workflows:

Intelligence stageLLM agentic analogKey LLM-side failure mode
DirectionSystem prompt / user requestAnchoring from prompt framing; over-narrow scoping
CollectionRetrieval / tool use / context assemblyPosition bias in long context (Liu 2023); RAG retrieval gaps
ProcessingFiltering, deduplication, chunkingChunk-boundary context loss (Roberts 2025); over-aggressive filtering
AnalysisReasoning step / chain-of-thought / agent decisionsSycophancy, hallucination, premature closure, confirmation bias
DisseminationOutput formatting / synthesis / handoffOverconfidence in stated conclusions; loss of caveats; persona artifacts in delivery

The mapping isn’t metaphorical — it is structural. Both pipelines:

  • Take an unstructured question and produce an analytic product
  • Compound errors across stages
  • Have most-attention paid to the analysis stage, but most-cause-of-failure distributed across all stages
  • Benefit from structural interventions (SATs) rather than asking individual stages to “try harder”

Why This Matters for the SAT-LLM Thesis

A single SAT intervention at the analysis stage cannot fix a flawed cycle. The wiki’s SAT Pipeline page describes how to compose SATs across stages; this concept page provides the underlying framework for why multi-stage composition is the right unit of work.

It also informs Bias Evaluations: judges need to be applied per-stage, not only to the final output, because failures at earlier stages produce traces that look reasonable at the analysis stage but were doomed at collection.


Sources

See Also