Executive summary
The right AI architecture is the least complex system that can satisfy the decision, authority, evidence, and recovery requirements of a named workflow. Deterministic code is usually the strongest baseline. Retrieval is appropriate when the main uncertainty is evidence discovery. One agent can support adaptive sequencing inside one authority envelope. Multiple agents require a measurable benefit from specialization, parallelism, challenge, or information separation. Swarm techniques are justified only when decentralized local coordination is intrinsic to the problem. The interactive selector below returns a starting hypothesis, not an automated design approval.
Decision relevance: Select a starting architecture and identify the evidence that must exist before increasing autonomy or coordination complexity.
Interactive starting hypothesis
Which architecture family should you test first?
The selector uses a small set of architecture questions. It does not inspect your system or approve a design.
Result
Complete the questions to see a starting hypothesis.
The decision remains human-owned.
Architecture family comparison
Deterministic workflow
Use when: The sequence, validations, and exception rules can be expressed clearly and the workflow benefits from predictable cost, latency, and recovery.
Avoid when: The task genuinely requires adaptive evidence gathering or planning across changing conditions.
Minimum evidence
- State machine or rules
- Representative tests
- Exception handling
- Rollback and reconciliation
Retrieval-augmented generation
Use when: The main uncertainty is finding and synthesizing permitted evidence while the retrieval path can remain mostly fixed.
Avoid when: The system must choose among materially different plans or external actions.
Minimum evidence
- Source and tenant boundaries
- Retrieval quality
- Groundedness and citation checks
- Freshness and retention
One bounded agent
Use when: The task needs adaptive sequencing or tool use but still has one coherent goal, state, authority envelope, and accountable owner.
Avoid when: Roles are genuinely independent, require parallel work, or must preserve information boundaries.
Minimum evidence
- Control loop
- Tool contracts
- Trajectory evaluation
- Budgets, stop conditions, and human authority
Multi-agent system
Use when: Specialization, parallelism, independent challenge, information separation, or dynamic allocation improves a measured outcome over the single-agent baseline.
Avoid when: The extra roles are decorative, correlated, or difficult to evaluate and recover.
Minimum evidence
- Topology and role distinctions
- Message contracts
- Shared-state rules
- Team-level benchmark and failure containment
Swarm-style coordination
Use when: Decentralized local rules, large-scale search or allocation, resilience to individual failure, and emergent adaptation are intrinsic to the problem.
Avoid when: A named accountable decision path or centralized policy approval is required for each consequential action.
Minimum evidence
- Local rules and feedback
- Emergence bounds
- Global invariants
- Containment, simulation, and independent stop control
The architecture decision starts with the consequence
Do not begin with a framework name or a desired number of agents. Begin with the business decision, the people affected, the cost of a wrong action, the evidence needed, and the authority that must remain with accountable humans.
NIST's AI RMF treats governance, context mapping, measurement, and management as connected lifecycle functions. That supports an architecture process in which the use boundary and evidence model are established before implementation complexity is selected. [S1]
Five architecture families
The five families in this guide are not exhaustive, and real systems often combine them. They are useful because each introduces a different control surface and burden of proof.
- Deterministic workflow for stable rules and explicit transitions
- RAG for bounded retrieval and evidence-supported generation
- One agent for adaptive sequencing inside one coherent authority envelope
- Multi-agent architecture for measured specialization, parallelism, challenge, or information separation
- Swarm-style coordination for decentralized local rules and emergent allocation or search
Gate questions before adding an agent
A team should be able to answer each question in plain language. An unclear answer is evidence that the process or authority model needs work before orchestration is expanded.
- Can the required sequence and exceptions be expressed deterministically?
- Is the main uncertainty evidence retrieval, or does the system need to choose materially different actions?
- Can every tool and external effect be bounded, authorized, verified, and reversed or reconciled?
- Does the task have one coherent owner and context, or are roles genuinely distinct?
- Will parallel or independent work improve a measured outcome after coordination cost is included?
- Does decentralization create resilience or scale that a supervisor cannot provide?
- Can the organization evaluate the complete trajectory and recover from partial execution?
Require a simpler baseline
Compare the proposed design with a deterministic workflow, a fixed RAG pipeline, and one strong agent. Without that baseline, additional agents can look productive because they produce more messages and artifacts rather than because they improve the decision.
Research on multi-agent failures and team underperformance reinforces the need for task-specific benchmarking rather than assuming that a larger team is better. [S2] [S3]
Treat operating cost as architecture evidence
Cost includes more than model tokens. Count tool calls, retrieval, messages, retries, state, observability, human review, incident handling, and the time required to understand a failed run. Agentic retrieval research also shows that added reasoning and retrieval loops can be materially more expensive and do not automatically outperform a well-designed simpler pipeline. [S4]
Evidence required before release
The selected family changes the depth of evidence, but every production design needs a named use boundary, architecture map, versioned dependencies, representative evaluation, security review, human authority, monitoring, rollback, and a recorded release decision.
Use the related AI System Release Gate to turn this architecture hypothesis into an evidence-backed release, conditional release, hold, or stop decision.
Research boundary
The selector is an educational aid. It does not inspect a system, certify safety, estimate return on investment, or replace architecture, security, legal, procurement, or operational review. A real decision must use the actual workflow, evidence, constraints, and accountable owners.
Sources
Sources support the linked statements and terminology. They do not certify a system, establish buyer intent, or convert this research into a formal assurance.
- Artificial Intelligence Risk Management FrameworkNIST · Accessed 2026-08-01
Government framework
- Why Do Multi-Agent LLM Systems Fail?arXiv · Accessed 2026-08-01
Research paper
- When Agent Teams Underperform Their Strongest MemberarXiv · Accessed 2026-08-01
Research paper
- Is Agentic RAG Worth It? An Experimental Comparison of RAG ApproachesarXiv · Accessed 2026-08-01
Research paper
- Towards a Science of AI Agent ReliabilityarXiv · Accessed 2026-08-01
Research paper