A design that separates deterministic discovery from model-based assessment. Code or APIs gather structured findings first, then the model interprets, ranks or explains them in a bounded second pass. This pattern reduces token use and improves auditability when the discovery work is enumerable.
What Two-Phase Agent Architecture Is Optimising For
Two-phase agent architecture is a control pattern for keeping the first pass deterministic and enumerable, then reserving the model for interpretation, prioritisation or explanation. That separation reduces ambiguity in the discovery stage and makes it easier to audit what the system actually saw before any model judgement was applied.
The practical value is not just efficiency. When the discovery layer is bounded to structured inputs, teams can reason about coverage, repeatability and failure modes without depending on a model to “find” facts that should have been mechanically collected in the first place.
How the Discovery Phase Should Be Designed
The first phase should behave like an evidence collector, not a thinker. It should query APIs, enumerate assets, extract fields, normalise records and preserve enough structure that the second pass can explain rather than invent. This is where the design earns its auditability: the system can show what was gathered, from where, and under what conditions.
That separation works best when the discovery step is predictable and complete for the chosen scope. If the underlying task is not enumerable, a two-phase pattern becomes weaker because the model may end up compensating for missing data instead of assessing a bounded set of findings.
How the Assessment Phase Adds Judgment Without Expanding Scope
The second phase is where the model ranks, summarises, correlates or narrates the collected findings. Its job is to add value to an already assembled dataset, not to widen the search space or silently fill gaps. In good implementations, the model is constrained by the discovered inputs and cannot substitute speculation for collection.
This is especially useful when the output must be reviewed by humans or compared across runs. A bounded assessment step makes it clearer whether the result changed because the evidence changed, or because the model changed its interpretation.
Where This Pattern Fits Best in Security Workflows
Two-phase agent architecture is most effective in workflows where discovery can be expressed as a repeatable query and the judgment layer is the real source of value. Common examples include asset inventories, findings triage, policy summarisation, alert enrichment and compliance-style review tasks.
It is less useful when the system must improvise throughout the workflow or when the “discovery” stage already depends heavily on model inference. The pattern works because it preserves a clean boundary between structured collection and bounded reasoning.
Risk and Threat Considerations
When the discovery phase is incomplete or poorly constrained, the architecture can create a false sense of assurance: the model may confidently assess a partial dataset and make the omission look like a legitimate conclusion. That is a governance risk as much as a technical one, because missing inputs can be harder to spot once they are wrapped in a polished assessment.
Failure mechanism: The system either skips records during deterministic collection or allows the model to treat incomplete findings as authoritative, so the second pass reinforces an error rather than exposing it.
Impact: Coverage gaps, missed high-risk items, inconsistent decisions and weak audit trails can follow, especially when the output is used for prioritisation, reporting or automated action.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST CSF 2.0, NIST SP 800-53 Rev 5, CIS Controls v8 and OWASP ASVS set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | ID.AM-01 — Asset Inventory | Two-phase discovery depends on enumerating the scoped inputs first. |
| Recommendation — Maintain an authoritative inventory so the first phase can collect findings from a complete asset set. | ||
| NIST SP 800-53 Rev 5 | AU-6 — Audit Record Review, Analysis, and Reporting | The pattern improves review of collected evidence and supports traceable assessment. |
| CM-8 — System Component Inventory | Enumerated discovery maps directly to component inventory before any second-pass judgment. | |
| Recommendation — Use audit-review procedures to validate what the discovery phase collected before model interpretation. Keep component inventories current so the discovery phase has a bounded, trustworthy source set. | ||
| CIS Controls v8 | CIS-8 — Audit Log Management | Deterministic collection and later assessment both benefit from preserved, reviewable logs. |
| Recommendation — Centralise and retain logs so the assessment phase can be audited against the original findings. | ||
| OWASP ASVS | V16 — Security Logging and Error Handling | The architecture relies on preserving evidence and handling collection failures cleanly. |
| Recommendation — Instrument collection and error paths so model assessment can be traced to the underlying evidence. | ||
Practitioner Guidance
Why practitioners should care: The architecture only improves control when the boundary between collection and interpretation is explicit. If teams cannot show what was gathered before model involvement, they lose most of the auditability benefit that makes the pattern attractive in the first place.
What to watch for: Review whether the discovery step is complete, enumerable and reproducible enough that a model is truly assessing evidence rather than rescuing a vague workflow. Where that is not true, tighten the collection contract before relying on the second pass for decisions.
Related resources from NHI Mgmt Group
- Why do AI agent and ML pipeline identities complicate zero trust architecture?
- What do teams get wrong about multi-agent architecture selection?
- How can teams tell whether an agent architecture is actually working?
- How should organisations respond when trusted developer tooling exposes agent architecture?
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on October 22, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org