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
| # | Stage | What happens | Common failure modes |
|---|---|---|---|
| 1 | Direction | Consumer sets the intelligence question / requirement | Narrow framing; ambiguous question; mismatched scope |
| 2 | Collection | Information is gathered from sources (HUMINT, SIGINT, IMINT, OSINT) | Single-source dependence; collection-strategy gaps; source-bias |
| 3 | Processing | Raw collection is validated, translated, structured | Filtering errors; mis-translation; misclassification |
| 4 | Analysis | Analysts extract insight, evaluate hypotheses, form judgments | Cognitive bias (see Cognitive Bias); mind-set lock-in; premature closure |
| 5 | Dissemination | Finished intelligence is communicated to decision-makers | Policymaker 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 stage | LLM agentic analog | Key LLM-side failure mode |
|---|---|---|
| Direction | System prompt / user request | Anchoring from prompt framing; over-narrow scoping |
| Collection | Retrieval / tool use / context assembly | Position bias in long context (Liu 2023); RAG retrieval gaps |
| Processing | Filtering, deduplication, chunking | Chunk-boundary context loss (Roberts 2025); over-aggressive filtering |
| Analysis | Reasoning step / chain-of-thought / agent decisions | Sycophancy, hallucination, premature closure, confirmation bias |
| Dissemination | Output formatting / synthesis / handoff | Overconfidence 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
- Grey Dynamics — Intelligence Failure (2024) — concise practitioner overview of the cycle with case studies (Iraq 2003, Russia/Ukraine 2022) at each stage
- CIA Tradecraft Primer (2009) — embeds the cycle implicitly throughout; SATs are positioned as interventions primarily at the analysis stage
See Also
- SAT Pipeline — how to apply SATs across the stages
- Bias Evaluations — per-stage measurement
- Cognitive Bias — the analysis-stage failure mode the wiki focuses on
- Structured Analytic Techniques — the toolkit applied at the analysis stage